Sea surface temperature forecast statistical correction method and system based on machine learning

Abstract

The invention relates to the technical field of temperature forecasting and machine learning, and provides a sea surface temperature forecasting statistical correction method based on machine learning, and the method comprises the following steps: obtaining sea surface temperature forecasting data of a target area and sea surface temperature observation data of the target area; constructing a statistical correction model; a neural network used for calculating model parameters of a statistical correction model is constructed, features reflecting the position of a target area and the sea surface temperature of the target area are input into the neural network, and the neural network outputs the model parameters of the statistical correction model; the sea surface temperature forecast data, the sea surface temperature observation data and the model parameters are input into the statistical correction model, and the statistical correction model calculates a sea surface temperature forecast statistical correction result based on the model parameters, the sea surface temperature forecast data and the sea surface temperature observation data; by adopting the method, the sea surface temperature forecasting error can be reduced, and the forecasting precision is improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of temperature prediction and machine learning, and particularly relates to a sea surface temperature prediction statistical correction method and system based on machine learning. BACKGROUND

[0002] Statistical correction is a key technology for reducing sea surface temperature prediction error and improving prediction accuracy. Global weather models can provide global sea surface temperature predictions, providing valuable reference for marine heat waves and fishery management. Although the original predictions generated by global weather models can provide a wealth of prediction information, they also contain complex systematic biases and random errors. Statistical correction is based on historical prediction and observation samples to build a model, aiming to eliminate the systematic bias of the original prediction and quantify the random error. The widely used statistical correction models include quantile mapping and joint Gaussian distribution model methods.

[0003] Traditional prediction statistical correction models are fitted based on samples from the same location and the same season to consider the spatiotemporal variability of predictions and observations. For example, for global sea surface temperature predictions in grid form, prediction and observation data for each grid are usually extracted for separate model fitting and prediction correction. The number of model fitting increases with the number of grids. Meanwhile, the fitted model cannot consider the prediction information of other grids, limiting the effectiveness of statistical correction. SUMMARY

[0004] Therefore, it is necessary to provide a sea surface temperature prediction statistical correction method based on machine learning to solve the above technical problems.

[0005] To solve the above technical problems, the technical solutions of the present application are as follows: A sea surface temperature prediction statistical correction method based on machine learning, comprising the following steps: obtaining sea surface temperature prediction data of a target area and sea surface temperature observation data of the target area; building a statistical correction model; building a neural network for calculating model parameters of the statistical correction model, inputting features reflecting the location of the target area and its sea surface temperature into the neural network, and outputting model parameters of the statistical correction model from the neural network; inputting the sea surface temperature prediction data, the sea surface temperature observation data, and the model parameters into the statistical correction model, and calculating a sea surface temperature prediction statistical correction result based on the model parameters, the sea surface temperature prediction data, and the sea surface temperature observation data of the statistical correction model.

[0006] Compared with the prior art, the technical solutions of the present application have the following advantages: The model parameters of the statistical correction model are calculated by using the neural network, automatic estimation of the parameters is realized, the influence of sample variability of the traditional parameter estimation on a small sample size is reduced, and the effectiveness of the parameter estimation is improved; and based on the model parameters, the sea surface temperature prediction data and the sea surface temperature observation data, the statistical correction result of the sea surface temperature prediction is calculated by using the statistical correction model, the sea surface temperature prediction error is reduced, and the prediction accuracy is improved. BRIEF DESCRIPTION OF DRAWINGS

[0007] Figure 1 A flowchart of the sea surface temperature prediction statistical correction method based on machine learning proposed in embodiment 1 is shown. Figure 2 An updated schematic diagram of the statistical correction model parameters based on machine learning error back propagation proposed in embodiment 2 is shown. Figure 3 A schematic diagram of the change process of the loss function value in model training proposed in embodiment 3 is shown. Figure 4 A CRPSS box plot of the original and corrected prediction proposed in embodiment 3 is shown. Figure 5 A schematic diagram of the proportion of the original and corrected prediction CRPSS greater than-5% proposed in embodiment 3 is shown. Figure 6 A schematic diagram of the original and corrected prediction case analysis proposed in embodiment 3 is shown. DETAILED DESCRIPTION

[0008] The drawings are only used for illustrative description, and cannot be understood as a limitation on the patent; In order to better illustrate the embodiments, some parts of the drawings may be omitted, enlarged or reduced, and do not represent the actual size of the product; It is understandable that some known structures and their descriptions in the drawings may be omitted for those skilled in the art.

[0009] The technical solutions of the present application will be further described below in combination with the drawings and embodiments.

[0010] Embodiment 1 The present embodiment proposes a sea surface temperature prediction statistical correction method based on machine learning, Figure 1 A flowchart of the sea surface temperature prediction statistical correction method based on machine learning proposed in the present embodiment is shown.

[0011] As Figure 1 shown, the sea surface temperature prediction statistical correction method based on machine learning of the present embodiment includes the following steps: S1: obtaining the sea surface temperature prediction data of the target area and the sea surface temperature observation data of the target area; S2: constructing a statistical correction model; S3: constructing a neural network for calculating model parameters of a statistical correction model, inputting features reflecting a location of a target area and its sea surface temperature into the neural network, and outputting model parameters of the statistical correction model from the neural network; S4: inputting the sea surface temperature prediction data, the sea surface temperature observation data and the model parameters into the statistical correction model, and calculating a sea surface temperature prediction statistical correction result based on the model parameters, the sea surface temperature prediction data and the sea surface temperature observation data of the statistical correction model. In an optional embodiment, expressions of the sea surface temperature prediction data of the target area and the sea surface temperature observation data of the target area include:

[0012] In the expressions, represents a sea surface temperature prediction data set of the target area; is at least composed of a prediction time, a prediction period, a collection member, a latitude and a longitude, a prediction value of 5 dimensions, a short-hand notation of ; represents a sea surface temperature observation data set of the target area; is at least composed of a target time, a latitude and a longitude, an observation value of 3 dimensions, a short-hand notation of ; in time, the target time is matched with the prediction and observation data by equating the target time to the sum of the prediction time and the prediction period, i.e. .

[0013] In an optional embodiment, the statistical correction model is a joint Gaussian distribution model or a mathematical model based on quantile mapping method.

[0014] In an optional embodiment, when the statistical correction model is a joint Gaussian distribution model, the statistical correction model represents a correlation between the prediction and the observation by a bivariate joint Gaussian distribution, and an expression of the correlation includes:

[0015] In the expression, represents a collection prediction mean corresponding to the sea surface temperature prediction data set, and respectively represent a mean and a standard deviation corresponding to the sea surface temperature prediction data set, and ​​​​​​respectively represent the mean and standard deviation of the sea surface temperature observation dataset, represents the correlation coefficient of sea surface temperature prediction and observation; and , , , and are model parameters of the statistical correction model; Based on the expression of the correlation relationship, the sea surface temperature prediction statistical correction result is represented as the conditional probability distribution of the observation given the prediction:

[0016] In the formula, represents the ensemble prediction mean corresponding to the corrected ensemble prediction, also represents the sea surface temperature prediction statistical correction result, is a normal distribution with as the mean and as the variance.

[0017] In an optional embodiment, when the statistical correction model is a mathematical model based on quantile mapping method, the statistical correction model corrects the prediction by matching the cumulative probability distribution of the prediction and the observation edge distribution to obtain the sea surface temperature prediction statistical correction result, and the expression of the sea surface temperature prediction statistical correction result includes:

[0018]

[0019] In the formula, represents the cumulative probability distribution of the obtained sea surface temperature prediction of the target area, represents the cumulative probability distribution of the edge distribution of the obtained sea surface temperature observation of the target area, and here a normal distribution is adopted, so the cumulative probability distribution is characterized by the mean and standard deviation parameters; represents the ensemble member of the sea surface temperature prediction data, represents the corresponding prediction value, represents the corresponding sea surface temperature prediction statistical correction result, represents the inverse function of .

[0020] In an optional embodiment, the expression of the forward calculation process of the neural network includes:

[0021] In the formula, represents the forward calculation process of the neural network, represents the activation function, For the neural network The weight matrix of the layer, Represents the neural network The output of the layer, Represents the neural network Layer bias; The expressions of the model parameters output by the neural network include:

[0022] Where, Represents the scaled output parameter , Represents the Sigmoid function, whose output range is , and Represents parameters respectively The upper and lower bounds of represents any parameter in the statistical correction model.

[0023] In an optional embodiment, the accuracy of the statistical correction result is higher than the original sea surface temperature forecast; The expression for the accuracy of the statistical correction result of the sea surface temperature forecast includes:

[0024] Where, represents the accuracy of the statistical correction results of the sea surface temperature forecast, and is a percentage; represents the continuously ranked probability score calculated for the reference climatological state forecast obtained by fitting a normal distribution to a sample of historical sea surface temperature observations, It represents the continuous ranking probability score calculated by the obtained sea surface temperature forecast of the target area and its corresponding sea surface temperature forecast statistical correction result.

[0025] In an optional embodiment, the neural network is trained to obtain model parameters of the statistical correction model; and the forecast statistical correction is performed using the obtained model parameters; The steps of training the neural network include: S1: Obtain a plurality of retrospective sea surface temperature forecast sets for a target area and their corresponding features reflecting the location of the area and the sea surface temperature to form a training set; wherein the features reflecting the location of the target area and the sea surface temperature of the target area are selected from an optional feature set, wherein the optional feature set includes: the data grid longitude, latitude, multi-year mean and multi-year standard deviation of the forecast, and the multi-year mean and multi-year standard deviation of the observation for each forecast grid cell; and obtain a plurality of retrospective sea surface temperature ensemble forecast data for the target area and their corresponding sea surface temperature observation data; S2: inputting the training set into the neural network, the neural network outputting model parameters of the statistical correction model, an expression of the model parameters comprising:

[0026] wherein, represents a set of model parameters, represents time, t , a forecast period, l , a latitude, y , a longitude, x corresponding statistical correction model parameters; S3: inputting a plurality of sea surface temperature retrospective forecast data and the set of model parameters into the statistical correction model, the statistical correction model using the set of parameters to respectively calculate statistical correction results of each retrospective forecast set based on sea surface temperature forecast data of time, t , a forecast period, l , a latitude, y , and a longitude, x of each retrospective forecast set; S4: calculating a loss function value of the neural network based on the statistical correction results corresponding to each retrospective forecast set, sea surface temperature observation data of the time corresponding to the retrospective forecast, and a preset loss function, after obtaining the loss function value, performing error back propagation calculation, i.e., calculating partial derivatives of the loss function with respect to each parameter of the neural network based on the chain rule, thereby updating the parameters in the neural network. S5: in the training process, iteratively performing steps S2-S4 until the loss function value of the neural network reaches a minimum or the number of iterations reaches a preset value, stopping calculation, and obtaining the trained neural network.

[0027] In an optional embodiment, when the correction statistical model is a joint Gaussian distribution model, the expression of the loss function comprises:

[0028]

[0029]

[0030] wherein, represents a continuous ranking probability score calculated based on the Gaussian distribution, is a sea surface temperature observation value corresponding to the sea surface temperature retrospective forecast; , , , and They correspond to the correlation coefficient between forecast and observation, the mean and standard deviation of sea surface temperature observation, and the mean and standard deviation of sea surface temperature forecast, respectively. They are all model parameters output by the neural network during training, that is, 、 、 、 and All belong to ; represents the ensemble forecast mean of the original retrospective sea surface temperature forecasts; and represent the cumulative distribution function and probability density function of the standard normal distribution respectively; When the corrected statistical model is a mathematical model based on the quantile mapping method, the expression of the loss function includes:

[0031] Where, represents the continuous ranking probability score CRPS calculated based on the ensemble forecast, represents the total number of ensemble members in the retrospective forecast ensemble; and Respectively based on 、 、 and , calculated using the revised statistical model and The predicted value corresponding to each ensemble member; 、 、 and They correspond to the mean and standard deviation of the sea surface temperature observations and the mean and standard deviation of the sea surface temperature forecasts, respectively, and are all model parameters calculated by the neural network during the training process, that is, 、 、 and All belong to .

[0032] This embodiment also proposes a correction system for a statistical correction method of sea surface temperature forecast based on machine learning, including: A data acquisition module is used to obtain sea surface temperature forecast data and sea surface temperature observation data of the target area; a model parameter calculation module, on which the statistical correction model and the neural network are provided, for calculating the model parameters of the statistical correction model based on the neural network; The statistical correction result output module is used to enable the statistical correction model to calculate the statistical correction result of the sea surface temperature forecast based on its model parameters, sea surface temperature forecast data and sea surface temperature observation data.

[0033] Embodiment 2 Based on the machine learning based sea surface temperature prediction statistical correction method proposed in Embodiment 1, a specific implementation example of the method is proposed in this embodiment, Figure 2 The specific implementation example of the machine learning error back propagation based statistical correction model parameter updating method proposed in this embodiment is shown in the following.

[0034] As Figure 2 shown, this embodiment uses a machine learning algorithm to estimate the parameters of the statistical correction model on a global scale, improves the effectiveness of the sea surface temperature prediction statistical correction, and derives a differentiable form of the traditional statistical correction model. The calculation process is realized by Python programming language, so as to realize the updating of the parameters of the statistical correction model by error back propagation.

[0035] The specific implementation example is shown as follows: (1) Taking global sea surface temperature prediction and observation as input data, considering its high-dimensional characteristics, the commonly used data format is NetCDF. First, read the global sea surface temperature prediction and observation by using Python language;

[0036] In the formula, F represents the global sea surface temperature prediction data set, f the prediction value, a total of 5 dimensions, the reporting time s , the prediction period l , the ensemble member m , the latitude y and the longitude x ; O represents the global sea surface temperature observation data set, a total of 3 latitudes, the target time t , the latitude y and the longitude x . In time, the target time is equal to the sum of the reporting time and the prediction period to match the prediction and observation data, that is .

[0037] (2) Taking the joint Gaussian distribution model and the quantile mapping method as an example, the form of the statistical correction prediction is derived, and the specific steps are as follows: The joint Gaussian distribution model represents the correlation between the prediction and the observation by a bivariate joint Gaussian distribution:

[0038] In the formula, represents the ensemble prediction mean, , , , and are the 5 parameters of the joint Gaussian distribution model. Based on the joint Gaussian distribution model, the updated forecast can be represented as the conditional probability distribution of the observation given the forecast:

[0039] It can be seen that, given the model parameters, the updated ensemble forecast can be represented as a normal distribution with as the mean and as the variance.

[0040] The quantile mapping method performs forecast updating by matching the cumulative probability distribution (CDF) of the forecast and the observation marginal distribution:

[0041] where and represent the CDF of the original forecast and the observation marginal distribution, respectively. Here, the normal distribution is adopted, and the CDF is characterized by the mean and standard deviation parameters, i.e., , , and . The updated forecast is represented as:

[0042] where represents the inverse function of the observation marginal distribution CDF.

[0043] It can be seen that, given the model parameters, the updated forecasts of the joint Gaussian distribution model and the quantile mapping method can be obtained from equations (3) and (5), respectively.

[0044] (3) Construct a neural network, combine the updated forecast derived in (2), and update the parameters in the error backpropagation manner. The overall process is shown in Figure 2 , and the specific steps are as follows: S1. Construct a neural network Take the data grid longitude, latitude, and other features as input and the statistical updating model parameters as output to construct a neural network:

[0045] where represents the forward calculation process of the neural network, represents the activation function, is the weight matrix of the i layer, represents the output of the i −1 layer, represents the output of the iBias of layers. The input and output of the neural network are flexible. The input can select the data grid longitude, latitude, predicted multi-year average value and multi-year standard deviation, and observed multi-year average value and multi-year standard deviation, etc. to reflect the regional location and its sea surface temperature characteristics. The output of the neural network depends on the parameters of the statistical correction model. The Gaussian distribution model has 5 parameters, so 5 output variables are needed. The quantile mapping method needs 4 output variables. At the same time, the structure of the neural network can also be adjusted according to the needs, such as introducing convolutional neural network and recurrent neural network, etc. The activation function can be selected according to the effect.

[0046] The output layer of the neural network can be scaled according to the value range of the parameters:

[0047] In the formula, represents the scaled output parameter , represents the Sigmoid function, and the output value range is [0, 1], and represent the upper and lower bounds of the parameter, respectively. represents a parameter in the statistical correction model, for example, when is the parameter of the joint Gaussian distribution model , and can be set to -0.99 and 0.99, respectively, to meet the value range of the parameter . The scaling of the neural network output is to adjust the value range of the output parameter, considering the actual physical meaning of the parameter, which helps the neural network to give reasonable parameter values.

[0048] S2. Generating corrected forecasts Based on the parameters of the statistical correction model output by the neural network, the corrected forecasts of the retrospective forecasts (past forecasts) can be directly generated.

[0049] For the joint Gaussian distribution model, the corrected forecasts can be directly calculated based on the parameters output by the neural network:

[0050] In the formula, represents the j th retrospective forecast set mean corresponding to the corrected forecast, , , , and represent the joint Gaussian distribution model parameters output by the neural network.

[0051] For quantile mapping, the revised forecast can be expressed as:

[0052] where, , , and denote the quantile mapping model parameters output by the neural network, denote the ensemble members of the j th retrospective forecast m , and its revised forecast.

[0053] Thus, the retrospective revised forecast is:

[0054] where, denote the retrospective revised forecast of the p th grid, and there are J ensemble forecasts, i.e. . The revised forecast can have different forms according to different statistical revision models: for the joint Gaussian distribution model, the revised forecast is a normal distribution with mean and variance ; the quantile mapping method gives the revised forecast in the form of a finite number of ensemble members.

[0055] S3. Loss function calculation Based on the retrospective revised forecast generated in S2, the forecast evaluation index is calculated as the loss function of the neural network, and the verification of the ensemble forecast can use the continuous ranking probability score (CRPS).

[0056] For the joint Gaussian distribution model, the CRPS calculation method based on the Gaussian distribution is used:

[0057] where, denotes the CRPS calculated based on the Gaussian distribution, is the sea surface temperature observation corresponding to the retrospective forecast, and are the mean and variance of the conditional probability distribution calculated based on the neural network output, i.e. , ; and denote the cumulative distribution function and probability density function of the standard normal distribution, respectively.

[0058] For the quantile mapping method, the CRPS calculation method based on ensemble forecast is adopted:

[0059] Where, represents the CRPS calculated based on the ensemble forecast, M Indicates the number of set members.

[0060] S4. Parameter update based on error back propagation As can be seen from S1, S2, and S3, the forward computational process of the neural network can be summarized as follows: the neural network outputs parameters, a statistical correction model is constructed based on these parameters, the correction model is used to perform retrospective forecast corrections, and the loss function is then calculated using the retrospective corrected forecasts. After obtaining the loss function value, the error is backpropagated, that is, the partial derivatives of the loss function with respect to each neural network parameter are calculated stepwise based on the chain rule, thereby updating the neural network parameters. Data from different longitude and latitude grids are extracted and the computational process S1 to S3 is repeated repeatedly to update the neural network parameters, thereby improving the effectiveness of the statistical correction model parameters generated by the neural network. As can be seen, the neural network is trained based on global sea surface temperature forecasts and observations. As the neural network trains, the statistical correction model parameters are continuously updated.

[0061] (4) Based on the neural network trained in (3), the statistical correction model parameters of each grid cell are output, thereby completing the correction of the global sea surface temperature forecast and carrying out the forecast. The specific steps are as follows: S1. Statistical Correction Model Parameter Set Based on the neural network trained in (3), the parameters of each grid cell in the world are:

[0062] Where, represents the global parameter set, Indicates time is t The forecast period is l , latitude is y , longitude is x The statistical correction model parameter set of , for the joint Gaussian distribution model, Include 、 、 、 and , for the quantile mapping method, Include 、 、 and .

[0063] S2. Statistical Revision of Forecasts For the joint Gaussian distribution model, the revised prediction is:

[0064] where, is the original prediction set mean is the corresponding revised prediction, , , , and denote the single grid, time parameter of the output of the trained neural network.

[0065] For the quantile mapping method, the revised prediction is:

[0066] where, denotes the original set prediction member .

[0067] S3. Prediction evaluation The revised prediction evaluates the prediction accuracy by calculating the CRPS as shown in equations (11) and (12). Further, the continuous ranking probability skill score (CRPSS) can be calculated to evaluate the relative performance of the revised prediction and the reference prediction:

[0068] where, denotes the CRPS calculated by the reference climatological prediction, the reference climatological prediction is obtained by fitting the historical observation sample to a normal distribution, denotes the CRPS calculated by the original and revised prediction.

[0069] The above method is used to carry out sea surface temperature prediction revision, and the advantages are: The application constructs a sea surface temperature prediction statistical revision method based on machine learning, which helps to reduce the global meteorological model sea surface temperature prediction error and improve the prediction accuracy. The application estimates the prediction statistical revision model parameters by using the machine learning method, can fully utilize the information of the global data set, reduce the sample variability influence of the traditional parameter estimation when the sample size is small, and improve the effectiveness of parameter estimation. The application makes the parameter estimation more flexible through the neural network, can select different input features and neural network structures, realizes automatic estimation of parameters, and is convenient and stable.

[0070] Embodiment 3 This embodiment is based on the sea surface temperature prediction statistical revision method based on machine learning proposed in embodiment 1, and proposes a specific implementation example, Figure 3The change process of the loss function value in the model training proposed in this embodiment is shown in the schematic diagram, Figure 4 The original and revised forecast CRPSS box plot proposed in this embodiment, Figure 5 The original and revised forecast CRPSS greater than -5% ratio diagram proposed in this embodiment, Figure 6 The original and revised forecast case analysis diagram proposed in this embodiment.

[0071] As Figures 3-6 shown, the present embodiment constructs a sea surface temperature forecast statistical revision method based on machine learning, which is mainly used for revising sea surface temperature forecast and improving forecast accuracy. The sea surface temperature forecast used in this example is derived from National Centers for Environmental Prediction's Climate Forecast System version 2 (NCEP-CFSv2), and the observation dataset is Optimal Interpolation sea surface temperature version 2 (OISST V2) sea surface temperature observation product. The revision and verification of global sea surface temperature forecast are carried out, and the specific implementation steps are as follows: S1. Read data: This step extracts sea surface temperature forecast and observation data. The NCEP-CFSv2 forecast and OISST V2 observation dataset used in this example are both grid data stored in NetCDF. The data is read through open_mfdataset and open_dataset in Python third-party library Xarray. The target time extracted is from 1983 to July 2024, and the forecast lead time is 0 months; S2. Model construction: This step is based on the mathematical modeling process described in the present invention, including neural network, statistical revision model, CRPS-based loss function calculation and error back propagation calculation, which are encapsulated into functions. The code implementation is mainly through Python third-party library PyTorch to realize automatic gradient calculation; S3. Model training: This step is based on the data read in S1 and the model constructed in S2, and the model is trained. The global sea surface temperature forecast and observation are used to update the weights and biases of the neural network, and improve the effectiveness of the statistical revision model parameters; S4. Revision forecast generation and verification: This step uses the neural network trained in S3 to output statistical revision model parameters for different grids, revises the future sea surface temperature forecast, and calculates the forecast verification index to verify the forecast accuracy.

[0072] The purpose of this embodiment is to revise the global sea surface temperature forecast. The specific implementation steps of this embodiment are as follows: S1. Read OISST V2 sea surface temperature observations using the open_dataset function in Xarray and store them in the da_obs variable, and read NCEP-CFSv2 sea surface temperature forecasts using open_mfdataset and store them in the da_fcst variable, where the data from 1983 to 2010 is used for model training, and the data from 2011 to 2024 is considered as real-time prediction for prediction correction.

[0073] S2. Based on the machine learning-based prediction statistical correction method proposed in the present application, mathematical modeling and code packaging are carried out, and the main implementation steps are as follows: S21: Based on the traditional statistical correction model, the prediction correction process based on machine learning is derived, mainly including the mathematical modeling process of equations (8) and (9); S22: Define the neural network structure, the input features include the latitude and longitude of the data grid, the correlation coefficient of the original prediction and observation in the training sample, the multi-year average and standard deviation of the sea surface temperature observation, the multi-year average and standard deviation of the original prediction of the sea surface temperature, the full connection neural network structure is adopted, the size of a single hidden layer is 128, and the size of the output layer is the same as the number of parameters of the statistical correction model; S23: Based on the definition of CRPS, realize the CRPS calculation process based on Gaussian distribution and ensemble prediction, mainly based on equations (11) and (12); S24: Program to realize the above model calculation process, and package it as a Python class, the Python class of joint Gaussian distribution is DplJointGaussian, the Python class of quantile mapping method is DplQuantileMapping, the loss function is CRPSLoss, and it is saved as a.py file.

[0074] S3. Based on the data read in S1 and the model developed in S2, model training is carried out, and the specific steps are as follows: S31: Based on the data read in S1, calculate the correlation coefficient, mean and variance, etc. as the input of neural network, store it as ds_dataset variable, regard the data of each grid as 1 sample, construct PyTorch DataLoader object to facilitate the extraction of different slice data from global dataset; S32: Perform rolling model training, specifically, when correcting the prediction of 2011, use the data from 1983 to 2010 as training data; when correcting the prediction of 2011, use the data from 1984 to 2011 as training data, and so on, the length of the training sample is fixed at 28 years; S33: Set the number of cycles Epoch to 500, in each cycle, call the DplJointGaussian class to get the retrospective corrected forecast, call the CRPSLoss to calculate the CRPS of the retrospective corrected forecast as the loss function value, and then update the parameters, and the optimizer is Adam of PyTorch; S34: Record the loss function value obtained in each round of training, and draw the average loss function value with the change of cycles by Matplotlib, Figure 3 The loss function value is shown with 1983-2010 data as the training sample. It can be seen that for quantile mapping and joint Gaussian distribution, the loss function value decreases with the increase of the number of model training, indicating the improvement of the effectiveness of the statistical correction model parameters, and finally the loss function value stabilizes at about 0.2; S35: After the neural network completes the training, output the statistical correction model parameters on each grid, save them in the ds_param variable, and output and save them as a NetCDF file.

[0075] S4. Extract the parameters obtained by training in S3, correct the forecast, and evaluate the performance of the forecast correction. The main implementation steps are as follows: S41: Based on the observation data in da_obs, calculate the CRPS of the original forecast, the corrected forecast and the climatological reference forecast for each grid, and further calculate the CRPSS; S42: Extract the CRPSS of the original forecast and the quantile mapping and joint Gaussian distribution corrected forecast based on neural network, and draw the box plot by Seaborn. It can be seen that both correction models can effectively improve the forecast accuracy, and the joint Gaussian distribution has higher CRPSS and better correction performance; S43: Take -5% as the threshold, and if the CRPSS is greater than -5%, it is considered that the forecast is not worse than the climatological reference forecast. Draw a column chart by Seaborn to show the proportion of grids with CRPSS greater than -5% in the global ocean grid. It can be seen that the proportion of the original forecast is less than 80%, while the quantile mapping method and the joint Gaussian distribution based on neural network are 91% and 96% respectively, indicating that the global sea surface temperature presents higher consistency; A grid is selected for case study, in addition to the neural network-based statistical correction, a traditional joint Gaussian distribution model is also used, that is, only the data on the grid is used for model fitting, the sea surface temperature prediction from 2011 to July 2024 and the observation are drawn by Matplotlib, the [25%, 75%] and [10%, 90%] uncertainty intervals of the ensemble prediction are represented by dark and light columns, and the observation is represented by a color point, it can be seen that the original prediction has a certain random error, although the traditional joint Gaussian distribution model can improve the CRPSS to a certain extent, the quantile mapping method based on the neural network and the joint Gaussian distribution show better prediction improvement performance, which is mainly due to the neural network in the global scale for statistical correction model parameter estimation, which can fully utilize the information from different regions, and improve the effectiveness of the model parameters.

[0076] Obviously, the above embodiments of the present application are only examples for clearly illustrating the present application, and are not intended to limit the embodiments of the present application. Based on the above description, other different forms of changes or variations can be made by those skilled in the art. Here, all the embodiments are not required to be exhausted. Any modification, equivalent replacement and improvement made within the spirit and principle of the present application shall be included in the protection scope of the claims of the present application.

Claims

1. A statistical correction method for sea surface temperature forecast based on machine learning, characterized in that: The following steps are involved: Obtain sea surface temperature forecast data and sea surface temperature observation data for the target area; Construct statistical correction models; Constructing a neural network for calculating model parameters of a statistical correction model, inputting features reflecting the location of the target area and its sea surface temperature into the neural network, and the neural network outputting the model parameters of the statistical correction model; The sea surface temperature forecast data, the sea surface temperature observation data and the model parameters are input into the statistical correction model, and the statistical correction model calculates the statistical correction result of the sea surface temperature forecast based on its model parameters, the sea surface temperature forecast data and the sea surface temperature observation data.

2. The statistical correction method for sea surface temperature forecast based on machine learning according to claim 1, characterized in that: The expressions of the sea surface temperature forecast data and the sea surface temperature observation data of the target area include: Where, represents the sea surface temperature forecast dataset for the target area; At least include the starting time , Forecast period , set members ,latitude and longitude The forecast values ​​of these five dimensions are The abbreviation symbol is ; A sea surface temperature observation dataset representing the target area; To include at least the target time ,latitude and longitude The observation values ​​of these three dimensions are The abbreviation symbol is In terms of time, the target time is equal to the sum of the starting time and the forecast period to match the forecast and observation data, that is, .

3. The statistical correction method for sea surface temperature forecast based on machine learning according to claim 1, characterized in that: The revised statistical model is a joint Gaussian distribution model or a mathematical model based on a quantile mapping method.

4. The statistical correction method for sea surface temperature forecast based on machine learning according to claim 3, characterized in that: When the statistical correction model is a joint Gaussian distribution model, the statistical correction model represents the correlation between the forecast and the observation through a two-variable joint Gaussian distribution, and the expression of the correlation includes: Where, represents the ensemble forecast mean corresponding to the sea surface temperature forecast dataset, and represent the mean and standard deviation of the sea surface temperature forecast dataset, respectively. and represent the mean and standard deviation of the sea surface temperature observation dataset, respectively. represents the correlation coefficient between the sea surface temperature forecast and observation; and 、 、 、 and are all model parameters of the statistical correction model; Based on the expression of the above correlation relationship, the statistical correction result of the sea surface temperature forecast is expressed as the conditional probability distribution of the observation given the forecast: Where, represents the ensemble forecast mean The corresponding corrected ensemble forecast also represents the statistical correction result of the sea surface temperature forecast. For is the mean, and is a normal distribution with a variance of .

5. The statistical correction method for sea surface temperature forecast based on machine learning according to claim 3, characterized in that: When the statistical correction model is a mathematical model based on the quantile mapping method, the statistical correction model performs forecast correction by matching the cumulative probability distribution of the forecast and the observed marginal distribution to obtain a statistical correction result of the sea surface temperature forecast. The expression of the statistical correction result of the sea surface temperature forecast includes: Where, represents the cumulative probability distribution of the sea surface temperature forecast for the target area, represents the cumulative probability distribution of the marginal distribution of the sea surface temperature observations obtained in the target area. Here, the normal distribution is adopted, and the cumulative probability distribution is characterized by the mean and standard deviation parameters; Represents the set member of the sea surface temperature forecast data The corresponding forecast value is express The corresponding statistical correction results of sea surface temperature forecast are: express The inverse function of .

6. The statistical correction method for sea surface temperature forecast based on machine learning according to claim 1, characterized in that: The expression of the forward calculation process of the neural network includes: Where, Represents the forward calculation process of the neural network, represents the activation function, For the neural network The weight matrix of the layer, Represents the neural network The output of the layer, Represents the neural network Layer bias; The expressions of the model parameters output by the neural network include: Where, Represents the scaled output parameter , Represents the Sigmoid function, whose output range is , and Represents parameters respectively The upper and lower bounds of represents any parameter in the statistical correction model.

7. The statistical correction method for sea surface temperature forecast based on machine learning according to claim 1, characterized in that: The accuracy of the statistical correction results is higher than that of the original sea surface temperature forecast; The expression for the accuracy of the statistical correction result of the sea surface temperature forecast includes: Where, represents the accuracy of the statistical correction results of the sea surface temperature forecast, and is a percentage; represents the continuously ranked probability score calculated for the reference climatological state forecast obtained by fitting a normal distribution to a sample of historical sea surface temperature observations, It represents the continuous ranking probability score calculated by the obtained sea surface temperature forecast of the target area and its corresponding sea surface temperature forecast statistical correction result.

8. The statistical correction method for sea surface temperature forecast based on machine learning according to any one of claims 1 to 7, characterized in that: Training the neural network to obtain model parameters of the statistical correction model; Use the obtained model parameters to carry out statistical revision of forecasts; The steps of training the neural network include: S1: Obtain a plurality of retrospective sea surface temperature forecast sets for a target area and their corresponding features reflecting the location of the area and the sea surface temperature to form a training set; wherein the features reflecting the location of the target area and the sea surface temperature of the target area are selected from an optional feature set, wherein the optional feature set includes: the data grid longitude, latitude, multi-year mean and multi-year standard deviation of the forecast, and multi-year mean and multi-year standard deviation of the observation; and obtain a plurality of retrospective sea surface temperature ensemble forecast data for the target area and their corresponding sea surface temperature observation data; S2: Inputting the training set into the neural network, the neural network outputs the model parameters of the statistical correction model, and the expression of the model parameters includes: Where, represents the set of model parameters, Indicates time is t The forecast period is l , latitude is y , longitude is x The corresponding statistical correction model parameter set; S3: Combine several sea surface temperature retrospective forecast data with model parameters Input statistical correction model, statistical correction model uses parameter set , based on the time of each retrospective forecast ensemble t The forecast period is l , latitude is y And the longitude is x The statistical correction results of each retrospective forecast ensemble are calculated based on the sea surface temperature forecast data. S4: Calculate the loss function value of the neural network based on the statistical correction results corresponding to each retrospective forecast set, the sea surface temperature observation data at the time corresponding to the retrospective forecast, and the preset loss function. After obtaining the loss function value, perform backpropagation calculation of the error, that is, gradually calculate the partial derivatives of the loss function with respect to each parameter of the neural network based on the chain rule, thereby updating the parameters in the neural network; S5: During the training process, steps S2 to S4 are iteratively executed until the loss function value of the neural network reaches the minimum or the number of iterations reaches a preset value, then the calculation is stopped to obtain a trained neural network.

9. The statistical correction method for sea surface temperature forecast based on machine learning according to claim 8, characterized in that: When the corrected statistical model is a joint Gaussian distribution model, the expression of the loss function includes: Where, represents the continuous sorting probability score calculated based on Gaussian distribution, is the sea surface temperature observation corresponding to the retrospective sea surface temperature forecast; 、 、 、 and They correspond to the correlation coefficient between forecast and observation, the mean and standard deviation of sea surface temperature observation, and the mean and standard deviation of sea surface temperature forecast, respectively. They are all model parameters output by the neural network during training, that is, 、 、 、 and All belong to ; represents the ensemble forecast mean of the original retrospective sea surface temperature forecasts; and represent the cumulative distribution function and probability density function of the standard normal distribution respectively; When the corrected statistical model is a mathematical model based on the quantile mapping method, the expression of the loss function includes: Where, represents the continuous ranking probability score CRPS calculated based on the ensemble forecast, represents the total number of ensemble members in the retrospective forecast ensemble; and Respectively based on 、 、 and , calculated using the revised statistical model and The predicted value corresponding to each ensemble member; 、 、 and They correspond to the mean and standard deviation of the sea surface temperature observations and the mean and standard deviation of the sea surface temperature forecasts, respectively, and are all model parameters calculated by the neural network during the training process, that is, 、 、 and All belong to .

10. A correction system for a statistical correction method of sea surface temperature forecast based on machine learning, characterized in that: include: A data acquisition module is used to obtain sea surface temperature forecast data and sea surface temperature observation data of the target area; a model parameter calculation module, on which the statistical correction model and the neural network are provided, for calculating the model parameters of the statistical correction model based on the neural network; The statistical correction result output module is used to enable the statistical correction model to calculate the statistical correction result of the sea surface temperature forecast based on its model parameters, sea surface temperature forecast data and sea surface temperature observation data.

Images