Ocean heat wave interpretability machine learning prediction method based on multi-source data

Abstract

The invention provides an ocean heat wave interpretability machine learning prediction method based on multi-source data, and the method comprises the steps: collecting simulation data and observation data of a plurality of climate mode members of CMIP6 in a preset time period, carrying out the preprocessing of the simulation data and observation data, and enabling the preprocessed simulation data and observation data to serve as a training set and a test set; constructing a machine learning model for predicting ocean heat waves, and training the machine learning model based on the training set to obtain a trained machine learning model; testing the trained machine learning model by using the test set to obtain a prediction result of the ocean heat wave, and identifying a physical factor with the highest influence degree on the prediction of the ocean heat wave through the interpretability of the trained machine learning model; the method can effectively overcome the problem of insufficient observation data samples of a traditional method, and significantly improves the accuracy of heat wave prediction. Meanwhile, the key functions of different meteorological variables in the prediction process can be determined, so that the scientific interpretability and practicability of the model are improved.

Description

Technical Field

[0001] The present invention relates to the technical field of ocean heat wave prediction, and more specifically, to an interpretable machine learning prediction method for ocean heat waves based on multi-source data. Background Art

[0002] A marine heatwave (MHW) is a discrete, abnormally high-temperature ocean event typically characterized by its duration, intensity, frequency, and spatial distribution. In recent years, driven by global warming, the frequency of marine heatwaves has increased significantly worldwide, particularly in the ecologically sensitive South China Sea. These frequent occurrences have had varying degrees of impact on coral reef ecosystems, aquaculture, and coastal economies. Therefore, improving the ability to predict marine heatwaves is crucial for ecological protection and resource allocation.

[0003] Current marine heatwave prediction methods can be primarily categorized into dynamical models and data-driven approaches. Dynamical models rely on a detailed description of initial fields, boundary conditions, and air-sea processes. While they possess a solid physical foundation, they are highly sensitive to initial errors, computationally expensive, and limited in their ability to depict complex air-sea coupling. Data-driven approaches, on the other hand, focus on extracting statistical patterns from observational data and can, to a certain extent, capture nonlinear relationships between meteorological variables. However, their prediction accuracy is highly dependent on the number and representativeness of training samples. Summary of the Invention

[0004] To overcome the shortcomings of the above-mentioned existing marine heatwave prediction methods in the prior art, such as limited sample size, insufficient representativeness, and weak generalization ability, the present invention provides an interpretable machine learning prediction method for marine heatwaves based on multi-source data. CMIP6 (the sixth coupled model intercomparison project), ERA5 (global atmospheric reanalysis data), and NOAA (national open meteorological data) data are used as the data basis for machine learning. By combining multi-source big data with interpretable models, it can not only improve the accuracy of medium- and short-term predictions, but also further reveal the key physical mechanisms affecting the evolution of marine heatwaves.

[0005] In order to solve the above technical problems, the technical solutions of the present invention are as follows: A method for predicting marine heatwaves using interpretable machine learning based on multi-source data includes the following steps: S1: Collect simulated data and observation data from multiple CMIP6 climate model members within a preset time period, preprocess them separately, and use the preprocessed simulated data and observation data as training and test sets respectively; S2: Constructing a machine learning model for predicting marine heat waves, training the machine learning model based on the training set, and obtaining a trained machine learning model; S3: Using the test set to test the trained machine learning model to obtain the prediction results of the ocean heat wave, and identifying the physical factors that have the greatest impact on the ocean heat wave prediction through the interpretability of the trained machine learning model.

[0006] Preferably, in step S1, the observation data are specifically ERA5 data and / or NOAA data.

[0007] Preferably, in step S1, simulation data of multiple CMIP6 climate model members in a specific ocean area and a preset time period are collected, and the simulation data at least includes any one or more of sea surface temperature data, wind field data, 2-meter air temperature data, and atmospheric pressure data; ERA5 data and NOAA data are collected in a specific ocean area and within a preset time period as the observation data; the ERA5 data include at least one or more of wind field data, 2-meter air temperature data, and atmospheric pressure data; and the NOAA data include at least sea surface temperature data and marine heat wave event observation results.

[0008] Preferably, in step S1, the preprocessing includes: uniformly aligning the time and spatial resolutions of the simulated data and the observed data, and then performing resampling, Z-score normalization, and combining time context information based on a sliding window, respectively, to obtain daily feature sequences corresponding to the simulated data and the observed data; Using an empirical orthogonal function to perform feature dimensionality reduction on the daily feature sequence, extracting a number of key mode time coefficient sequences; performing lag analysis on the time coefficient sequence of each key mode, and obtaining a time series feature sample containing lag information; The time series feature samples containing lag information corresponding to the simulated data and the observed data are used as the training set and the test set respectively.

[0009] Preferably, the adopting of an empirical orthogonal function to perform feature dimensionality reduction on the daily feature sequence comprises: EOF decomposition is performed on various meteorological data in the daily feature sequences of the simulation data and the observation data, and the first several modes that can represent the spatial variation pattern of a specific ocean area are extracted as the key modes. The time coefficient sequence corresponding to each key mode is obtained to achieve feature dimensionality reduction; The hysteresis analysis includes: For each of the key mode's time coefficient sequences, combined with the time tags in the marine heatwave event observations, a lag analysis is performed. The impact of different meteorological data on the occurrence of marine heatwaves at various lag times is evaluated, and the most significant time lag window is determined. The time coefficient sequence of the key mode and its most significant time lag window are saved together as the time series feature sample containing lag information.

[0010] Preferably, in step S2, the machine learning model includes at least one or more of a random forest model, a long short-term memory network, a multilayer perceptron, a support vector machine and an extreme gradient boosting model.

[0011] Preferably, in step S2, the machine learning model is specifically a random forest model; In step S3, the physical factors that have the greatest impact on marine heat wave prediction are identified through the interpretability of the trained machine learning model, including: Based on the trained random forest model, all decision trees are traversed to obtain the average information gain score of each type of meteorological data in all decision trees; at the same time, the Gini coefficient of each type of meteorological data is calculated and ranked; based on the average information gain score and Gini coefficient, key variables with global predictive influence are selected; The SHAP method is used to calculate the marginal Shapley value of each key variable in a single sample prediction, and a SHAP dependency diagram of each key variable is plotted in a visual manner; the SHAP dependency diagram is analyzed to obtain the key variables with the highest contribution to marine heat wave prediction, and the physical factors corresponding to the key variables are further determined to be the physical factors with the greatest impact on marine heat wave prediction.

[0012] Preferably, the method further comprises: revealing the nonlinear interaction relationship between variables by constructing a SHAP interaction diagram between different key variables, and evaluating the synergistic enhancement effect of different key variable combinations on marine heatwave prediction.

[0013] Preferably, the method further comprises: OOB out-of-bag error perturbation analysis is performed on the selected key variables to further improve the stability of variable importance assessment.

[0014] The present invention also provides a computer-readable storage medium having a computer program stored thereon, wherein the computer program implements the steps in the above method when executed by a processor.

[0015] Compared with the prior art, the beneficial effects of the technical solution of the present invention are: The present invention provides an interpretable machine learning prediction method for ocean heat waves based on multi-source data. First, simulation data and observation data of multiple CMIP6 climate model members within a preset time period are collected and preprocessed respectively. The preprocessed simulation data and observation data are used as a training set and a test set, respectively. Then, a machine learning model for predicting ocean heat waves is constructed, and the machine learning model is trained based on the training set to obtain a trained machine learning model. Finally, the trained machine learning model is tested using the test set to obtain ocean heat wave prediction results. The physical factors with the greatest impact on ocean heat wave prediction are identified through the interpretability of the trained machine learning model.

[0016] The present invention uses CMIP6 data, ERA5 data and NOAA data as the data basis for machine learning, which can systematically expand the number of training samples and climate background diversity of machine learning models, and solve the problems of traditional statistical methods being limited by sample size, insufficient training, and weak generalization ability. The present invention relies on long-term simulation data of multiple climate model members to construct a large-scale training set across models, years and spatial grids, and uses reanalysis data and satellite observation data for unified format processing and standardized verification, so that the model can capture the evolution laws and triggering mechanisms of marine heat waves under different climate scenarios, especially in short-, medium- and long-term (7-day, 30-day, 90-day) prediction scenarios. High-precision rolling predictions are achieved, thereby effectively improving the stability, accuracy and adaptability of the model under multiple time periods.

[0017] In addition, the present invention also introduces the global and local interpretability analysis mechanism of the RF model, and establishes a complete explanation system covering variable importance ranking, sample-level contribution evaluation, variable interaction response and time heterogeneity, to solve the problem of lack of model comprehensibility caused by the "black box" of existing deep learning models; the present invention uses Gini information gain, OOB, SHAP value and other means to comprehensively quantify the predictive influence of input variables, and identify the core role of variables such as SST, wind field, and air pressure under different prediction time limits; at the same time, combined with variable perturbation tests and false positive / missing sample diagnosis, the source and mechanism of model prediction errors are effectively tracked; the nonlinear influence relationship of key variables in time series and spatial grid points is displayed through SHAP heat maps and dependency diagrams, providing users with decision transparency and signal traceability, significantly enhancing the scientific support and credibility of the model in actual business applications. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] Figure 1 This is a flow chart of the explainable machine learning prediction method for marine heat waves based on multi-source data provided in Example 1.

[0019] Figure 2This is a schematic diagram of an interpretable machine learning prediction method for marine heat waves based on multi-source data provided in Example 1.

[0020] Figure 3 This is a comparison chart of the prediction accuracy of the three methods provided in Example 1 for the South China Sea marine heat wave with 7-day, 30-day and 90-day prediction periods.

[0021] Figure 4 This is a comparison chart of the predicted results of the three methods provided in Example 1 with the actual South China Sea marine heat wave at 7-day, 30-day and 90-day prediction periods.

[0022] Figure 5 This is a single variable importance analysis diagram for the random forest model provided in Example 2.

[0023] Figure 6 This is the SHAP feature importance distribution diagram of the random forest model provided in Example 2.

[0024] Figure 7 This is a SHAP waterfall plot of the random forest model provided in Example 2.

[0025] Figure 8 This is the SHAP two-factor dependency diagram of the random forest model prediction results (7-day prediction duration) provided in Example 2. DETAILED DESCRIPTION

[0026] The accompanying drawings are for illustrative purposes only and are not to be construed as limiting the present application; In order to better illustrate this embodiment, some parts in the drawings may be omitted, enlarged, or reduced, and do not represent the actual product size; It is understandable to those skilled in the art that some well-known structures and their descriptions may be omitted in the drawings.

[0027] The technical solution of the present invention is further described below with reference to the accompanying drawings and embodiments.

[0028] Example 1 like Figure 1 As shown, this embodiment provides an interpretable machine learning prediction method for marine heat waves based on multi-source data, including the following steps: S1: Collect simulated data and observation data from multiple CMIP6 climate model members within a preset time period, preprocess them separately, and use the preprocessed simulated data and observation data as training and test sets respectively; S2: Constructing a machine learning model for predicting marine heat waves, training the machine learning model based on the training set, and obtaining a trained machine learning model; S3: Testing the trained machine learning model using the test set to obtain a prediction result of the ocean heat wave, and identifying the physical factors that have the greatest impact on the ocean heat wave prediction through the interpretability of the trained machine learning model; In step S1, the observation data is specifically ERA5 data and / or NOAA data; In step S1, simulation data of multiple CMIP6 climate model members in a specific ocean area and a preset time period are collected, wherein the simulation data includes at least one or more of sea surface temperature data, wind field data, 2-meter air temperature data, and atmospheric pressure data; Collecting ERA5 data and NOAA data within a specific ocean area and a preset time period as the observation data; the ERA5 data includes at least one or more of wind field data, 2-meter air temperature data, and atmospheric pressure data; and the NOAA data includes at least sea surface temperature data and marine heat wave event observation results; In step S1, the preprocessing includes: uniformly aligning the time and spatial resolutions of the simulated data and the observed data, and then performing resampling, Z-score normalization, and combining temporal context information based on a sliding window, respectively, to obtain daily feature sequences corresponding to the simulated data and the observed data; Using an empirical orthogonal function to perform feature dimensionality reduction on the daily feature sequence, extracting a number of key mode time coefficient sequences; performing lag analysis on the time coefficient sequence of each key mode, and obtaining a time series feature sample containing lag information; The time series feature samples containing lag information corresponding to the simulated data and the observed data are used as the training set and the test set respectively; The adopting of an empirical orthogonal function to perform feature dimension reduction on the daily feature sequence includes: EOF decomposition is performed on various meteorological data in the daily feature sequences of the simulation data and the observation data, and the first several modes that can represent the spatial variation pattern of a specific ocean area are extracted as the key modes. The time coefficient sequence corresponding to each key mode is obtained to achieve feature dimensionality reduction; The hysteresis analysis includes: For each of the key mode's time coefficient sequences, combined with the time tags in the marine heatwave event observations, a lag analysis is performed. The impact of different meteorological data on the occurrence of marine heatwaves at various lag times is evaluated, and the most significant time lag window is determined. Saving the time coefficient sequence of the key mode and its most significant time lag window together as the time series feature sample containing lag information; In step S2, the machine learning model includes at least one or more of a random forest model, a long short-term memory network, a multilayer perceptron, a support vector machine, and an extreme gradient boosting model.

[0029] In the specific implementation process, Figure 2 Figure 1 shows a schematic diagram of the marine heatwave prediction method of this embodiment. First, a training set is constructed using data from multiple CMIP6 climate model members (such as ACCESS and BCC), key variables (SST, PSL, TAS, UAS, and VAS) are extracted, and unified preprocessing (such as cropping, resampling, and standardization) is used to generate training samples suitable for multi-timeframe (7-day, 30-day, and 90-day) predictions. After model training is completed, ERA5 and NOAA observational data are introduced as a test set, maintaining a consistent data processing process, and rolling predictions are performed. By comparing the prediction results with marine heatwave events identified in observations, the model's performance in terms of hit rate, timeliness, and spatial matching is evaluated, thereby verifying the model's generalization ability and practicality. Specifically, multiple CMIP6 model members, including mainstream models such as ACCESS, BCC, MRI, and CAN, were systematically selected. Daily-scale data covering the South China Sea region from 1950 to 2014 were collected to construct a large-scale simulation sample set across time and models. In this embodiment, the collected data included multiple meteorological variables such as sea surface temperature (SST), wind field (UAS, VAS), 2-meter air temperature (T2M), and atmospheric pressure (PSL), ensuring coverage of thermal and dynamic influencing factors and improving the integrity and diversity of model input features. The above data were then uniformly resampled and cropped in the South China Sea region (latitude 0°N to 30°N, longitude 100°E to 130°E), retaining ocean grid points and excluding land. ERA5 reanalysis data were then collected for the same region, mainly including UAS, VAS, T2M, and PSL. NOAA observation data were also collected, mainly including SST and marine heatwave event identification results, for verification and evaluation of model prediction results. In this embodiment, preprocessing includes: applying Z-score normalization technology to process variables on a grid-by-grid basis to enhance variable distribution consistency and alleviate model training convergence issues under numerical differences; constructing a daily feature sequence, combining temporal context information through a sliding window method, with each sample covering continuous historical variables, to predict whether a marine heatwave will occur on the target day; using the marine heatwave criterion defined by Hobday et al., generating a heatwave label based on the 90th percentile temperature threshold and the rule of at least five days of duration, forming a multidimensional supervisory signal; The empirical orthogonal function (EOF) decomposition method is used to extract the main spatial modal characteristics; the lag analysis method is used to evaluate the lag response of various meteorological variables to marine heat waves, and the optimal lag period is determined as the time window for subsequent model input; based on the above processing results, a unified format of data input samples is constructed for subsequent model training and prediction processing; specifically, EOF decomposition is first performed on multiple meteorological variables in the CMIP6 multi-member ensemble data and NOAA and ERA5 observation data. This process decomposes the original three-dimensional space-time field data based on the covariance matrix, extracts the first few modes that can represent the main spatial change patterns of the South China Sea and its surrounding areas, and obtains the corresponding time coefficient sequence, thereby achieving high The authors then perform feature compression and structural extraction of dimensional meteorological variables. For each of the aforementioned time coefficient sequences, combined with the time labels of observed marine heatwave events, they perform daily lagged correlation tests. By systematically evaluating the impact of different meteorological variables (such as SST, VAS, and UAS) on the occurrence of MHWs at various lag times, they identify the most significant time lag windows. The core goal of this lag analysis is to identify climate precursor signals that may appear before MHWs and extract key predictive factors that precede heatwave outbreaks in time. Ultimately, these time series feature variables containing lagged information are constructed as inputs to a machine learning model, giving the model forward-looking physical guidance capabilities and improving the scientific rationality and accuracy of the predictions. To support multi-time prediction, this implementation introduces multiple lag windows, such as 7, 30, and 90 days, and combines historical input sequences to construct multi-task training objectives for the short, medium, and long term. Grid search and five-fold cross-validation are used to optimize the hyperparameters of five models: random forest (RF), long short-term memory (LSTM), multilayer perceptron (MLP), support vector machine (SVM), and extreme gradient boosting (XGBoost), to build optimal models for different structures. In this embodiment, the XGBoost model uses 1000 trees with a maximum tree depth of 10, a learning rate of 0.1, and a row and column sampling ratio of 0.8 to improve model accuracy while effectively suppressing overfitting; the RF model consists of 1000 decision trees with a maximum depth of 10, and each leaf node contains at least 5 samples; the SVR model uses a radial basis kernel function, the regularization parameter C is set to 1.0, and the kernel function coefficient is automatically determined by the feature dimension; the MLP model consists of two hidden layers, each containing 64 neurons, using a ReLU activation function and a mean square error loss function, and using Dropout regularization and sliding window input features during training. The number of training rounds is 100, the optimizer is Adam, and the learning rate is 0.001; the LSTM model consists of three unidirectional LSTM hidden layers, each containing 128 units, using a batch size of 64, using an MSE loss function, and training on the standardized sliding window data for 100 rounds. Dropout regularization is used to suppress overfitting during training; After model training, ERA5 (2015-2022) and NOAA (2015-2022) observational data were used as test data, maintaining the same data processing process as CMIP6. The trained model was used to make rolling predictions for ERA5 and NOAA, predicting marine heatwaves in the South China Sea for the next 7, 30, and 90 days for each date. Finally, the predicted results were matched with marine heatwaves identified in actual observations, and the model predictions were compared with the actual observations to analyze and contrast the model's effectiveness. Testing with actual observational data verified the model's generalization ability and stability, providing technical support for subsequent heatwave forecasting applications in the South China Sea and similar sea areas. In this embodiment, in order to more intuitively evaluate the prediction effect of the model, the prediction results of each model are compared with different methods; Figure 3 As shown, Figure 3 Comparison of the prediction accuracy of the three methods provided in this example for the South China Sea marine heatwave at 7-day, 30-day, and 90-day forecast periods; the black columns in the figure represent the prediction accuracy of the ERA5 reanalysis data, the blue columns represent the prediction accuracy of the three CMIP6 model members (ACCESS, BBC, and CAN), and the red columns represent the prediction accuracy of the five machine learning models (RF, XGB, LSTM, SVR, and MLP); the value on each column represents the prediction accuracy of the corresponding model; Figure 4 As shown, Figure 4 The three methods provided in this embodiment are compared with the prediction results of the models at 7-day, 30-day, and 90-day prediction periods and the actual South China Sea marine heatwave. Specifically, the scatter plot shows the occurrence of marine heatwaves and the prediction results of different machine learning models. The top row represents the actual occurrence of marine heatwaves, while the subsequent rows represent the predicted marine heatwaves of each machine learning model. Figure 3 and Figure 4 It can be seen that the method provided in this embodiment can have a higher prediction accuracy; In terms of model training and prediction, this method uses CMIP6 multi-member ensemble data as the main source of training samples, systematically collecting long-term series data in the South China Sea region from multiple model members such as ACC and ESS, covering a variety of key meteorological variables such as SST, wind speed, and air pressure, and combining them with ERA5 and NOAA observation data for independent testing and rolling verification. Through processing methods such as unified spatial resolution, variable standardization, and sliding window construction of historical series, a multi-task modeling system covering short-term (7 days), medium-term (30 days), and long-term (90 days) prediction scenarios is established. The model framework adopts multiple algorithms such as random forest, XGBoost, and LSTM, and performs parameter optimization, while integrating multi-member samples, significantly enhancing the model's generalization and multi-source adaptability, ensuring stable prediction performance under different climate backgrounds.

[0030] Example 2 This embodiment provides an interpretable machine learning prediction method for marine heat waves based on multi-source data, including the following steps: S1: Collect simulated data and observation data from multiple CMIP6 climate model members within a preset time period, preprocess them separately, and use the preprocessed simulated data and observation data as training and test sets respectively; S2: Constructing a machine learning model for predicting marine heat waves, training the machine learning model based on the training set, and obtaining a trained machine learning model; S3: Testing the trained machine learning model using the test set to obtain a prediction result of the ocean heat wave, and identifying the physical factors that have the greatest impact on the ocean heat wave prediction through the interpretability of the trained machine learning model; In step S1, the observation data is specifically ERA5 data and / or NOAA data; In step S1, simulation data of multiple CMIP6 climate model members in a specific ocean area and a preset time period are collected, wherein the simulation data includes at least one or more of sea surface temperature data, wind field data, 2-meter air temperature data, and atmospheric pressure data; Collecting ERA5 data and NOAA data within a specific ocean area and a preset time period as the observation data; the ERA5 data includes at least one or more of wind field data, 2-meter air temperature data, and atmospheric pressure data; and the NOAA data includes at least sea surface temperature data and marine heat wave event observation results; In step S1, the preprocessing includes: uniformly aligning the time and spatial resolutions of the simulated data and the observed data, and then performing resampling, Z-score normalization, and combining temporal context information based on a sliding window, respectively, to obtain daily feature sequences corresponding to the simulated data and the observed data; Using an empirical orthogonal function to perform feature dimensionality reduction on the daily feature sequence, extracting a number of key mode time coefficient sequences; performing lag analysis on the time coefficient sequence of each key mode, and obtaining a time series feature sample containing lag information; The time series feature samples containing lag information corresponding to the simulated data and the observed data are used as the training set and the test set respectively; The adopting of an empirical orthogonal function to perform feature dimension reduction on the daily feature sequence includes: EOF decomposition is performed on various meteorological data in the daily feature sequences of the simulation data and the observation data, and the first several modes that can represent the spatial variation pattern of a specific ocean area are extracted as the key modes. The time coefficient sequence corresponding to each key mode is obtained to achieve feature dimensionality reduction; The hysteresis analysis includes: For each of the key mode's time coefficient sequences, combined with the time tags in the marine heatwave event observations, a lag analysis is performed. The impact of different meteorological data on the occurrence of marine heatwaves at various lag times is evaluated, and the most significant time lag window is determined. Saving the time coefficient sequence of the key mode and its most significant time lag window together as the time series feature sample containing lag information; In step S2, the machine learning model is specifically a random forest model; In step S3, the physical factors that have the greatest impact on marine heat wave prediction are identified through the interpretability of the trained machine learning model, including: Based on the trained random forest model, all decision trees are traversed to obtain the average information gain score of each type of meteorological data in all decision trees; at the same time, the Gini coefficient of each type of meteorological data is calculated and ranked; based on the average information gain score and Gini coefficient, key variables with global predictive influence are selected; Perform OOB out-of-bag error perturbation analysis on each selected key variable to further improve the stability of variable importance assessment; The SHAP method is used to calculate the marginal Shapley value of each of the key variables in a single sample prediction, and a SHAP dependency diagram of each of the key variables is plotted in a visual manner; the SHAP dependency diagram is analyzed to obtain the key variables that contribute most to the marine heat wave prediction, and the physical factors corresponding to the key variables are further determined to be the physical factors that have the greatest impact on the marine heat wave prediction; By constructing SHAP interaction diagrams between different key variables, the nonlinear interaction relationship between variables is revealed, and the synergistic enhancement effect of different key variable combinations on marine heatwave prediction is evaluated.

[0031] In the specific implementation process, the method in this embodiment is consistent with the overall process of the method in Example 1. The difference is that the machine learning model used in this embodiment is fixed to a random forest model. The random forest model is used as the explanatory analysis object. The relative contribution of each meteorological feature to the forecast result is evaluated in combination with the interpretable analysis method under three typical forecast lead times of 7 days, 30 days, and 90 days. Specifically, a trained random forest (RF) model was constructed, and the average information gain score of each variable in all decision trees was obtained through the natural splitting path of the tree structure. The Gini coefficient calculation method was introduced to rank the purity contribution of all input features, preliminarily screening key variables with global predictive influence. At the same time, the out-of-bag error (OOB) was used to evaluate the marginal contribution of each variable to the generalization performance of the model, further improving the stability of variable importance assessment. The SHAP method was then introduced to interpret the prediction results of a single sample from a local perspective, revealing the positive and negative effects and interactions of different variables in the model decision-making process. In this step, the SHAP method was used to calculate the marginal Shapley value of each variable in the single sample prediction and establish a sample-level explanatory map. SHAPDependence plots were visually drawn to analyze how changes in variable values ​​nonlinearly affect the probability of marine heatwave predictions. SHAP ranking plots were drawn for different prediction timeframes (7 days, 30 days, and 90 days) to identify changes in the dominant factors of short-, medium-, and long-term models. By integrating global and local interpretation results, we identify key meteorological features that have a high contribution across different forecast horizons, identify physical factors that significantly influence marine heatwave forecasts, such as SST, UAS, and VAS, and explore their physical mechanisms, providing a basis for improving the scientific interpretability of the model. In addition, the SHAP interaction value plot can be used to reveal nonlinear interactions between variables, such as the synergistic enhancement effect of the combination of high SST and wind speed on the occurrence of marine heat waves; In this embodiment, in order to more intuitively analyze and clarify the contribution rate of each meteorological feature to the forecast results, and to identify the physical factors that have a key impact on the forecast of marine heat waves, the analysis results of this embodiment are analyzed by Figures 5 to 8 to conduct a display; Figure 5 The figure shows the importance of individual variables in the random forest model. The figure shows the importance of each predictor variable in the RF model used to predict marine heat waves on different days. Figure 6 The SHAP feature importance distribution map of the RF model is presented, showing the contribution and direction of the 15 most important predictive variables to the occurrence of marine heatwaves in the South China Sea under different forecast horizons (7 days, 30 days, and 90 days); Figure 7 A SHAP waterfall plot of the random forest model is shown, which shows the contribution and direction of the top ten important predictors when predicting the 100th and 1000th test samples at different prediction times; Figure 8 The SHAP two-factor dependency diagram of the random forest model prediction results (7-day prediction period) is displayed, showing the interaction between the six groups of predictors in the South China Sea marine heatwave prediction. In each sub-plot, the horizontal axis represents the actual value of the dominant variable in the group of variables, and the vertical axis represents the SHAP value corresponding to the dominant variable (i.e., the degree of contribution to the prediction result). Each data point represents a sample, and its color indicates the value of the second variable: red indicates a higher value of the second variable, and blue indicates a lower value of the second variable. In terms of interpretability analysis, this method introduces feature importance ranking and SHAP methods to analyze the influence of variables in the trained model and construct a global and local two-level explanation system; by ranking the importance of variables such as SST, UAS, VAS, and PSL, their key role mechanisms under different prediction time periods are revealed; at the same time, SHAP values ​​are used to achieve sample-level causal tracing, quantify the marginal contribution of each meteorological variable in the specific prediction results, and analyze the synergistic influence relationship between variables through interactive dependency diagrams; in addition, this method also visualizes SHAP values ​​as heat maps at the spatial level, identifies the distribution pattern of the influence intensity of variables in different sea areas, provides a basis for studying the formation process of marine heat waves and model prediction deviations, and significantly enhances the comprehensibility and application credibility of the prediction system.

[0032] The same or similar reference numerals correspond to the same or similar components; The terms used in the drawings to describe positional relationships are for illustrative purposes only and are not to be construed as limiting the present application. Obviously, the above embodiments of the present invention are merely examples for the purpose of clearly illustrating the present invention, and are not intended to limit the embodiments of the present invention. Those skilled in the art will appreciate that other variations or modifications can be made based on the above description. It is not necessary and impossible to enumerate all embodiments here. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention shall be included within the scope of protection of the claims of the present invention.

Claims

1. A method for predicting marine heat waves using interpretable machine learning based on multi-source data, characterized in that: The following steps are involved: S1: Collect simulated data and observation data from multiple CMIP6 climate model members within a preset time period, preprocess them separately, and use the preprocessed simulated data and observation data as training and test sets respectively; S2: Constructing a machine learning model for predicting marine heat waves, training the machine learning model based on the training set, and obtaining a trained machine learning model; S3: Using the test set to test the trained machine learning model to obtain the prediction results of the ocean heat wave, and identifying the physical factors that have the greatest impact on the ocean heat wave prediction through the interpretability of the trained machine learning model.

2. The method for predicting marine heatwaves using explainable machine learning based on multi-source data according to claim 1, characterized in that: In step S1, the observation data is specifically ERA5 data and / or NOAA data.

3. The method for predicting marine heatwaves using explainable machine learning based on multi-source data according to claim 2, characterized in that: In step S1, simulation data of multiple CMIP6 climate model members in a specific ocean area and a preset time period are collected, wherein the simulation data includes at least one or more of sea surface temperature data, wind field data, 2-meter air temperature data, and atmospheric pressure data; ERA5 data and NOAA data are collected in a specific ocean area and within a preset time period as the observation data; the ERA5 data include at least one or more of wind field data, 2-meter air temperature data, and atmospheric pressure data; and the NOAA data include at least sea surface temperature data and marine heat wave event observation results.

4. The method for predicting marine heatwaves using explainable machine learning based on multi-source data according to claim 3, characterized in that: In step S1, the preprocessing includes: uniformly aligning the time and spatial resolutions of the simulated data and the observed data, and then performing resampling, Z-score normalization, and combining temporal context information based on a sliding window, respectively, to obtain daily feature sequences corresponding to the simulated data and the observed data; Using an empirical orthogonal function to perform feature dimensionality reduction on the daily feature sequence, extracting a number of key mode time coefficient sequences; performing lag analysis on the time coefficient sequence of each key mode, and obtaining a time series feature sample containing lag information; The time series feature samples containing lag information corresponding to the simulated data and the observed data are used as the training set and the test set respectively.

5. The method for predicting marine heatwaves using explainable machine learning based on multi-source data according to claim 4, characterized in that: The adopting of an empirical orthogonal function to perform feature dimension reduction on the daily feature sequence includes: EOF decomposition is performed on various meteorological data in the daily feature sequences of the simulation data and the observation data, and the first several modes that can represent the spatial variation pattern of a specific ocean area are extracted as the key modes. The time coefficient sequence corresponding to each key mode is obtained to achieve feature dimensionality reduction; The hysteresis analysis includes: For each of the key mode's time coefficient sequences, combined with the time tags in the marine heatwave event observations, a lag analysis is performed. The impact of different meteorological data on the occurrence of marine heatwaves at various lag times is evaluated, and the most significant time lag window is determined. The time coefficient sequence of the key mode and its most significant time lag window are saved together as the time series feature sample containing lag information.

6. The method for predicting marine heatwaves using interpretable machine learning based on multi-source data according to any one of claims 1 to 5, characterized in that: In step S2, the machine learning model includes at least one or more of a random forest model, a long short-term memory network, a multi-layer perceptron, a support vector machine, and an extreme gradient boosting model.

7. The method for predicting marine heatwaves using explainable machine learning based on multi-source data according to claim 6, characterized in that: In step S2, the machine learning model is specifically a random forest model; In step S3, the physical factors that have the greatest impact on marine heat wave prediction are identified through the interpretability of the trained machine learning model, including: Based on the trained random forest model, all decision trees are traversed to obtain the average information gain score of each type of meteorological data in all decision trees; at the same time, the Gini coefficient of each type of meteorological data is calculated and ranked; based on the average information gain score and Gini coefficient, key variables with global predictive influence are selected; The SHAP method is used to calculate the marginal Shapley value of each key variable in a single sample prediction, and a SHAP dependency diagram of each key variable is plotted in a visual manner; the SHAP dependency diagram is analyzed to obtain the key variables with the highest contribution to marine heat wave prediction, and the physical factors corresponding to the key variables are further determined to be the physical factors with the greatest impact on marine heat wave prediction.

8. The method for predicting marine heatwaves using explainable machine learning based on multi-source data according to claim 7, characterized in that: The method also includes: revealing the nonlinear interaction relationship between variables by constructing a SHAP interaction diagram between different key variables, and evaluating the synergistic enhancement effect of different key variable combinations on marine heatwave prediction.

9. The method for predicting marine heatwaves using explainable machine learning based on multi-source data according to claim 7, characterized in that: The method further comprises: OOB out-of-bag error perturbation analysis is performed on the selected key variables to further improve the stability of variable importance assessment.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the method according to any one of claims 1 to 9 are implemented.

Images