Marine environment forecasting method integrating day-by-day climate state and machine learning model

Abstract

The invention discloses a day-by-day climate state and machine learning model fused marine environment forecasting method. The method mainly comprises the steps of determining a forecasting target as a marine environment parameter, a spatial position and a time range; collecting and processing marine environment parameter data, and performing quality control and cleaning; constructing a marine environment parameter data set and a day-by-day climate state data set obtained by averaging years of marine environment parameter historical data; the method comprises the following steps: building a machine learning model by taking marine environment parameter anomaly as a forecasting object, and training and optimizing the built machine learning model; performing reverse normalization processing on a marine environment parameter anomaly forecasting result output by the optimized machine learning model; and adding the data after the reverse normalization processing and the day-by-day climate state data to obtain a marine environment forecasting result of the target area. The marine environment forecasting method provides thoughts for development of future environment forecasting, and is expected to provide more accurate and stable services for the fields of resource management, ecological protection, coastal engineering and the like.

Description

Technical Field

[0001] The present invention relates to the research field of marine environment forecasting, and in particular to integrating the daily climate state of marine environmental parameters with a machine learning model for the first time to improve the effect of marine environment forecasting. Background Art

[0002] Marine environmental forecasting is crucial for offshore production, environmental protection, marine economy, disaster response, and scientific research. In past studies, daily climate states have been used to predict climate change [1]. Daily climate states are statistical data based on historical meteorological or oceanographic elements, and they generally perform well in predicting climate change, especially when climate conditions are relatively stable. However, the marine environment is a complex process controlled by multiple factors. Its changes are not only affected by meteorological factors such as solar radiation, atmospheric temperature, and wind speed, but also by the interaction of earth system elements such as ocean surface characteristics, ocean currents, and topography. The small and medium-scale dynamic processes within it cannot be ignored. This complex regulatory network makes the marine environment a highly dynamic nonlinear system, which limits the predictive effect of daily climate states. Numerical models are also an important means of marine environmental forecasting. They simulate the dynamic processes in the marine system through modeling of a set of physical equations to predict future changes in the marine environment. However, numerical models are usually affected by multiple factors, including complex topography, uncertainty in boundary conditions, and the approximation of parameterization schemes, and their prediction accuracy is limited to a certain extent.

[0003] In recent years, the application of machine learning technology in marine environmental forecasting has gradually increased. Machine learning models can learn the basic patterns and complex nonlinear relationships hidden in the data from massive observations and reanalysis data, and in some cases can produce more accurate prediction results than traditional numerical models. Previous researchers have achieved relatively accurate 7-30 day forecasts of regional marine environment based on machine learning models and various methods [2,3]. However, machine learning models have limitations: due to the limitation of input data length, they may not contain long-term information, and are prone to performance degradation in long-term predictions. In the current study, the daily climate state of marine environmental parameters is only used to compare the forecasting effects of machine learning and other methods. However, the daily climate state can capture the trend and periodicity of long-term climate change, and is stable and reliable for long-term forecasts. Therefore, combining the advantages of daily climate state with machine learning models is expected to improve the accuracy of long-term marine environmental forecasts. So far, no research has involved the integration of daily climate state with machine learning models to achieve the forecast of marine environmental parameters, and this method has not yet formed a complete system and implementation process.

[0004] In this context, by combining the daily climate state of marine environmental parameters with machine learning methods, the accuracy, reliability and long-term stability of marine environmental forecasts can be further improved based on the basic understanding of the physical processes of the system through the daily climate state and the learning ability of the machine learning model for complex relationships. In general, the marine environmental forecasting method that integrates the daily climate state and machine learning model proposed in the present invention achieves further improvement and enhancement of machine learning marine environmental forecasts, provides new ideas and methods for the development and application of marine environmental forecasting technology in the future, and can provide more accurate and stable forecasting services for fields such as marine resource management, ecological protection, marine engineering and climate change response.

[0005] [References]

[0006] [1]Hansen JW, Challinor A, Ines A, et al. Translating climate forecasts into agricultural terms: advances and challenges[J]. Climate research, 2006, 33(1):27-41.

[0007] [2]Shao Q, Hou G, Li W, et al. Ocean reanalysis data-driven deep learning forecast for sea surface multivariate in the South China Sea[J]. Earth and Space Science, 2021, 8(7):e2020EA001558.

[0008] [3]Xu S,Dai D,Cui X,et al.Adeep learning approach to predict seasurface temperature based on multiple modes[J].Ocean Modelling,2023,181:102158. Summary of the invention

[0009] In view of the above-mentioned prior art, the present invention provides a method for forecasting the marine environment by integrating the daily climate state with the machine learning model, which can better capture the periodic change trend and nonlinear characteristics in the marine environment, thereby improving the accuracy of machine learning forecasts. Through the marine environment forecasting experiment in a certain area, the long-term stable forecast of marine environmental parameters is achieved, providing basic support for more reliable marine environment forecasts.

[0010] In order to solve the above technical problems, the present invention proposes a method for forecasting the marine environment by integrating daily climate state with a machine learning model, comprising the following steps:

[0011] Step 1: Determine the forecast target: Determine the forecast target as the marine environmental parameters, as well as the spatial location and time range;

[0012] Step 2: Data collection and processing: Collect multi-year historical data of the forecast target described in step 1 through satellite observation and / or field observation, unify the temporal and spatial dimensions of data from different data sources, and perform quality control and cleaning; data processing includes removing outliers and filling missing values, and the method of removing outliers includes one or more of statistical methods, box plots, and empirical removal, and the means of filling missing values ​​includes one or more of mean or median filling, regression method, interpolation method, K-nearest neighbor algorithm, and random forest;

[0013] Step 3: Construct the marine environment parameter dataset and daily climate state dataset, including:

[0014] 3-1) Based on the data collected and processed in step 2, construct a marine environmental parameter data set;

[0015] 3-2) Based on the marine environmental parameter data set, a statistical method is used to average the marine environmental parameter data for many years to obtain daily climate state data and construct a daily climate state data set;

[0016] Step 4: Build a machine learning model with the anomaly of ocean environmental parameters as the forecast object, and train and optimize the built machine learning model in turn: including:

[0017] 4-1) Use one of the following methods to build a machine learning model: linear regression, decision tree, neural network model, or ensemble learning model:

[0018] 4-2) Generate input data for the machine learning model: subtract the corresponding daily climate state data from the daily data of the ocean environmental parameters to obtain the ocean environmental parameter anomaly as the target variable of the model; use the ocean environmental parameter anomaly data and the daily climate state data as prediction factors to generate the input feature matrix of the machine learning model; normalize the target variable and the input feature matrix to generate the input data of the machine learning model;

[0019] 4-3) Machine learning model training: The input data of the machine learning model is divided into a training set, a validation set, and a test set. During the training process, according to the relationship between the input feature matrix and the target variable and the set iteration number threshold, the mean square error (MSE) or root mean square error (RMSE) is used as the loss function, and the gradient descent method or Adam optimization algorithm is used to adjust the model parameters to minimize the prediction error;

[0020] 4-4) Machine learning model optimization: Use the validation set data to evaluate the trained model, and use one of the methods of grid search, random search and Bayesian optimization to optimize the model parameters according to the evaluation results to obtain the optimized machine learning model;

[0021] Step 5: Denormalize the marine environmental parameter anomaly forecast results output by the optimized machine learning model, add the denormalized data to the daily climate data, and obtain the marine environmental forecast results for the target area.

[0022] Furthermore, the marine environment forecasting method of the present invention comprises:

[0023] In step 1, the ocean environment parameters include one or more of ocean surface temperature, ocean salinity, sea surface height and ocean current velocity; the spatial position refers to the forecast sea area; and the time range refers to the forecast start time and duration.

[0024] In step 4-1), the neural network model includes one of a convolutional neural network (CNN), a long short-term memory neural network (LSTM) and a transformer model (Transformer).

[0025] In step 4-2), the anomalies of ocean environmental parameters are introduced as prediction factors based on dynamic mechanism analysis or existing experience.

[0026] In step 5, after obtaining the marine environment forecast results of the target area, the marine environment forecast results are visualized, including a time series graph and a spatial distribution graph.

[0027] Compared with the prior art, the present invention has the following beneficial effects:

[0028] Compared with the deficiencies in stability and accuracy of traditional marine environment forecasting methods, the marine environment forecasting method of the present invention, by integrating daily climate states and machine learning models, gives full play to the advantages of both, and improves the accuracy and reliability of machine learning marine environment forecasts. Daily climate states have the ability to capture long-term climate change trends and periodic signals, and can achieve stable and reliable long-term forecasts, while machine learning models can extract complex nonlinear relationships from massive data to improve the accuracy of forecasts. The marine environment forecasting method of the present invention brings new ideas to the development and application of future marine environment forecasting technology, and provides more reliable forecasting support for responding to marine environmental changes and related issues.

[0029] The marine environment forecasting method of the present invention can not only provide more accurate and stable marine environment forecasting services for fields such as marine resource management, ecological protection, marine engineering and climate change response, but also better support relevant decisions and actions to ensure the sustainable development of marine production and life, environmental protection and marine economy. BRIEF DESCRIPTION OF THE DRAWINGS

[0030] Figure 1 It is a flowchart of the marine environment forecasting method of the present invention;

[0031] Figure 2 It is a display of the sea surface temperature forecast results of the material research of the present invention, wherein: (a) is the spatial distribution of forecast error of the traditional machine learning method, and (b) is the spatial distribution of forecast error of the daily climate state and machine learning fusion method proposed in the present invention.

[0032] Figure 3 The material studies the changing relationship between the overall root mean square error of the area and the forecast time, and is used to present and compare the forecasting effect of the daily climate state and machine learning fusion method proposed in the present invention with that of the traditional machine learning method. DETAILED DESCRIPTION

[0033] The present invention will be further described below in conjunction with the accompanying drawings and specific material research, but the following material research is by no means limiting the present invention in any way.

[0034] This material study uses sea surface temperature forecast as an example to conduct experiments. The data source is the satellite time-delay quality control data of OSTIA (The Operational Sea Surface Temperature and Ice Analysis), which is a satellite sea surface temperature and sea ice analysis product developed and maintained by the UK Met Office. This product integrates observation data from multiple satellites, including AVHRR of the National Oceanic and Atmospheric Administration (NOAA), AATSR of the European Space Agency (ESA), MODIS and VIIRS, and combines numerical models and data assimilation technology to provide high-quality data on sea surface temperature and sea ice worldwide. In the material study, the sea surface temperature from January 1, 1993 to June 30, 2020 in the OSTIA dataset was selected as historical data to predict the sea surface temperature from July 1, 2020 to June 30, 2021.

[0035] like Figure 1 As shown, the marine environment forecasting method proposed in the present invention that integrates daily climate state and machine learning model includes the following steps, and each step in the present invention can be implemented by MATLAB software or Python programming language.

[0036] Step 1: Determine forecasting goals

[0037] In the present invention, the forecast target is determined to be the ocean environmental parameters, as well as the spatial position and time range; the ocean environmental parameters include one or more of the ocean surface temperature, ocean salinity, sea surface height and ocean current velocity; the spatial position refers to the forecast sea area; the time range refers to the start time and duration of the forecast.

[0038] In this research material, taking the sea surface temperature forecast as an example, the sea area with a latitude and longitude range of 5°N~40°N and 108°E~180°E is selected as the forecast area. Within this spatial range, the method of integrating the daily climate state and the machine learning model proposed in this invention is used to make 12 monthly scale forecasts of the sea surface temperature. The time resolution of the forecast is daily, and the forecast time range is from July 1, 2020 to June 30, 2021. Each monthly scale forecast inputs the data of the first 30 days to obtain the sea surface temperature forecast value for the next 30 days.

[0039] Step 2: Data collection and processing

[0040] In the present invention, many years of historical data of the forecast target described in step 1 are collected through satellite observation and / or field observation, the temporal and spatial dimensions of data from different data sources are unified, and quality control and cleaning are performed; data processing includes removing outliers and filling missing values, the method of removing outliers includes one or more of statistical methods, box plots and empirical removal, and the means of filling missing values ​​includes one or more of mean or median filling, regression method, interpolation method, K-nearest neighbor algorithm, and random forest;

[0041] In this study, the historical sea surface temperature data of the OSTIA dataset for many years were collected. The collected data were unified in time and space, with a time resolution of daily and a spatial resolution of 1°×1°. The data after the time and space dimensions were unified were quality controlled and cleaned, outliers in the data were removed by empirical removal methods, and missing values ​​were filled by linear interpolation methods.

[0042] Step 3: Construct ocean environment parameter dataset and daily climate state dataset

[0043] Step 3-1) constructs a sea surface temperature dataset based on the data collected and processed in step 2, with a latitude and longitude range of 5°N to 40°N and 108°E to 180°E. The spatial resolution is 1°×1° and the temporal resolution is daily.

[0044] Step 3-2) Based on the marine environmental parameter data set, a statistical method is used to average the marine environmental parameter data over many years to obtain daily climate state data and construct a daily climate state data set.

[0045] In this research material, based on the sea surface temperature dataset constructed in step 3-1, the multi-year sea surface temperature data from January 1, 1993 to December 31, 2016 are averaged daily to construct a daily climate state dataset. For example, the average value of all January 1s from 1993 to 2016 is taken as the data of January 1 of the daily climate state, and so on to obtain 366 days of daily climate state data.

[0046] Step 4: Build a machine learning model with the anomaly of ocean environmental parameters as the forecast object, and train and optimize the built machine learning model in turn

[0047] Step 4-1) Build a machine learning model

[0048] In the method of the present invention, a method among linear regression, decision tree, neural network model and integrated learning model can be used to build a machine learning model. The neural network model includes one of CNN, LSTM and Transformer. In this research material, CNN is used to build a machine learning model. Sea surface temperature data usually has spatial correlation, that is, sea areas in adjacent positions may have similar sea surface temperature distributions. The CNN model can effectively extract spatial features, identify local patterns and structures in the image through convolution operations and pooling operations, and thus capture the spatial information of sea surface temperature. The CNN machine learning model structure built in this material research includes three convolution layers, and batch normalization, ReLU activation function and average pooling layer are set after each convolution layer for feature extraction and downsampling. Finally, the Dropout layer is used to prevent overfitting, and the Flatten layer flattens the feature map into a one-dimensional vector, and finally the prediction result is output through the fully connected layer.

[0049] Step 4-2) Generate input data for the machine learning model

[0050] First, the ocean environment parameter anomaly is obtained by subtracting the corresponding daily climate state data value from the daily data of the ocean environment parameters, which is used as the target variable of the model; in this research material, the sea surface temperature anomaly data is used as the target variable of the model. Specifically, based on the constructed sea surface temperature data set and daily climate state data set, the corresponding daily climate state data value is subtracted from the sea surface temperature data to obtain the sea surface temperature anomaly data. The sea surface temperature anomaly refers to the temperature difference between the daily sea surface temperature and the daily climate state, which reflects the degree of deviation between the sea surface temperature and the long-term climate average. Then, the ocean environment parameter anomaly data and the daily climate state data are used as prediction factors to generate the input feature matrix of the machine learning model. In this research material, the sea surface temperature anomaly data selected by experience and the daily climate state data are used together as prediction factors to generate the input feature matrix of the machine learning model; the target variable and the input feature matrix are normalized to generate the input data of the machine learning model. The specific methods in this research material are as follows:

[0051] 1) Traverse the sea surface temperature in the forecast area spatially grid by grid. For each grid point in the area, input the sea surface temperature anomaly and daily climatological data around the point (2°×2°) as the feature matrix, and use the sea surface temperature anomaly of the point as the target variable;

[0052] 2) Normalize the target variables and input feature matrices to ensure that the data has similar scales and ranges, and improve the stability and convergence speed of model training. Standardize the mean and variance of each input feature matrix, calculate the mean and standard deviation of each input feature matrix, and then transform the values ​​of each input feature matrix as follows:

[0053]

[0054] Where x′ is the normalized eigenvalue, x is the original eigenvalue, μ is the mean of the feature, and σ is the standard deviation of the feature.

[0055] Step 4-3) Machine Learning Model Training

[0056] The input data of the machine learning model is divided into a training set, a validation set, and a test set according to the time range. The training set is from January 1, 2012 to December 31, 2016, the validation set is from January 1, 2017 to January 30, 2017, and the test set is from July 1, 2020 to June 30, 2021. The model is trained using the training set data, and the input feature matrix and the target variable are staggered by 30 days to achieve monthly forecasts. During the training process, the machine learning model adjusts the model parameters according to the relationship between the input feature matrix and the target variable to minimize the prediction error. In the present invention, mean square error or root mean square error can be used as the loss function, and the model parameters can be adjusted using the gradient descent method or the Adam optimization algorithm to minimize the prediction error. In this research material, the mean square error (MSE) is selected as the loss function, and the Adam optimizer is used to optimize the model parameters. According to the duration of the training data, the batch size is selected as 12, and the initial number of iterations is set to 150 times.

[0057] Step 4-4) Machine Learning Model Optimization

[0058] For sea surface temperature forecast, the Bayesian optimization hyperparameter optimization method is used to train the best model. First, the hyperparameter space is defined to ensure the best balance between model performance and generalization ability. The model performance index is used as the objective function. In this research material, the accuracy on the validation set is used as the optimization target, and the validation set data is used to evaluate the accuracy of the trained model, aiming to maximize the accuracy of the model. According to the evaluation results, Bayesian optimization is used to optimize the model parameters. The Bayesian optimizer selects the next hyperparameter combination that is most likely to improve performance based on the existing experimental results and model performance, and uses these parameters to train a new model. In each iteration, the performance of the model will be recorded for subsequent analysis and comparison. Finally, based on the performance of the objective function, the model parameter combination with the best performance is selected to obtain the optimized machine learning model.

[0059] Step 5: Obtain the marine environment forecast results for the target area

[0060] The marine environmental parameter anomaly forecast results output by the optimized machine learning model are denormalized, and the denormalized data are added to the daily climate data to obtain the marine environmental forecast results for the target area.

[0061] In this research material, the optimized machine learning model was used to make 12 monthly forecasts, and the sea surface temperature anomaly forecast results for a total of 360 days were output. The sea surface temperature anomaly was post-processed by reverse normalization to ensure that the forecast results had the same scale and range as the actual observed values. The reverse normalized data was added to the daily climate data to construct the sea surface temperature forecast results. The forecast results obtained were visualized, such as time series graphs, spatial distribution graphs, etc., to intuitively show the changing trend and spatial distribution of the forecast results.

[0062] In the research materials, the model performance is evaluated and verified. By comparing the model prediction results with the actual observation data, the prediction accuracy of the prediction model at different time and space scales is checked. The error analysis methods include but are not limited to the calculation of indicators such as root mean square error (RMSE), mean absolute error (MAE), and correlation coefficient to quantify the prediction accuracy of the model. Analyze the prediction accuracy and possible influencing factors to verify the effectiveness of the marine environmental parameter prediction method. The verified marine environmental forecast model will be released to relevant users or decision makers for production, life, emergency response and other related work, providing strong marine environmental forecast data support.

[0063] In this research material, after constructing the sea surface temperature forecast results, the forecast results are evaluated based on the root mean square error (RMSE) indicator. For the forecast RMSE corresponding to each spatial point, the following formula is used to solve:

[0064]

[0065] In the formula, the number of grid points in the target sea area is M, the forecast time span is N (360 days in this material study), the sea surface temperature forecast value is p, and the actual sea surface temperature is y. Each spatial point will generate a time series containing 360-day sea surface temperature forecast values, and the forecast RMSE of the spatial point can be calculated based on the sea temperature forecast value and the actual value.

[0066] The RMSE of the regional forecast of sea temperature for each day is calculated using the following formula:

[0067]

[0068] That is, the daily forecast RMSE is obtained by comparing the forecast value of each spatial point in the area with the true value.

[0069] Following the above steps, the daily climate state-machine learning model fusion method can be applied to predict sea surface temperature. Figure 2(b) is the SST forecast evaluation result obtained by the daily climate state-machine learning fusion method proposed in this invention. In the entire forecast area, the minimum RMSE is 0.22℃, the maximum is 2.93℃, and the RMSE of more than 74.3% of the spatial points is less than 1.00℃. In this material study, the forecast results obtained by the same steps of the traditional CNN machine learning forecast without the fusion of daily climate state are used as a comparison, such as Figure 2 As shown in (a), the minimum RMSE of each spatial point of the forecast results of the traditional machine learning method is 0.32℃, the maximum is 13.19℃, and only 45.6% of the spatial points have an RMSE less than 1.00℃. In general, the method combining daily climate state with machine learning has significant advantages over the traditional machine learning method, and the errors of more than 94.7% of the spatial points have been improved, especially in the coastal waters with complex terrain and marine environment, and the Kuroshio Sea area with fast flow, strong nonlinearity, and complex dynamic processes.

[0070] To further illustrate the advantages of the present invention, the change of RMSE of sea surface temperature forecast over time was analyzed in the material research. The RMSE time series of the traditional machine learning method and the daily climate state-machine learning fusion method are shown in Figure 2. Figure 3 The RMSE of traditional machine learning forecasts fluctuates between 1.07℃ and 2.47℃, while the forecast effect of the daily climate state-machine learning fusion method is stable, with a maximum error of only 1.14℃ and a minimum error of 0.69℃. The error of more than 92% of the days is less than 1℃. In summary, this method can significantly improve the forecast effect of sea surface temperature and further reduce the forecast error.

[0071] Although the present invention has been described above in conjunction with the accompanying drawings, the present invention is not limited to the above-mentioned specific embodiments, which are merely illustrative rather than restrictive. Under the guidance of the present invention, ordinary technicians in this field can make many improvements and changes without departing from the purpose of the present invention, which are all within the protection of the present invention.

Claims

1. A method for forecasting the marine environment by integrating daily climate states with a machine learning model, characterized in that: The steps include: Step 1: Determine the forecast target: Determine the forecast target as the marine environmental parameters, as well as the spatial location and time range; Step 2: Data collection and processing: Collect multi-year historical data of the forecast target described in step 1 through satellite observation and / or field observation, unify the temporal and spatial dimensions of data from different data sources, and perform quality control and cleaning; data processing includes removing outliers and filling missing values, and the method of removing outliers includes one or more of statistical methods, box plots, and empirical removal, and the means of filling missing values ​​includes one or more of mean or median filling, regression method, interpolation method, K-nearest neighbor algorithm, and random forest; Step 3: Construct the marine environment parameter dataset and daily climate state dataset, including: 3-1) Based on the data collected and processed in step 2, construct a marine environmental parameter data set; 3-2) Based on the marine environmental parameter data set, a statistical method is used to average the marine environmental parameter data for many years to obtain daily climate state data and construct a daily climate state data set; Step 4: Build a machine learning model with the anomaly of ocean environmental parameters as the forecast object, and train and optimize the built machine learning model in turn: including: 4-1) Use one of the following methods to build a machine learning model: linear regression, decision tree, neural network model, or ensemble learning model: 4-2) Generate input data for the machine learning model: subtract the corresponding daily climate state data from the daily data of the ocean environmental parameters to obtain the ocean environmental parameter anomaly as the target variable of the model; use the ocean environmental parameter anomaly data and the daily climate state data as prediction factors to generate the input feature matrix of the machine learning model; normalize the target variable and the input feature matrix to generate the input data of the machine learning model; 4-3) Machine learning model training: The input data of the machine learning model is divided into a training set, a validation set, and a test set. During the training process, according to the relationship between the input feature matrix and the target variable and the set iteration number threshold, the mean square error or root mean square error is used as the loss function, and the gradient descent method or Adam optimization algorithm is used to adjust the model parameters to minimize the prediction error; 4-4) Machine learning model optimization: Use the validation set data to evaluate the trained model, and use one of the methods of grid search, random search and Bayesian optimization to optimize the model parameters according to the evaluation results to obtain the optimized machine learning model; Step 5: Denormalize the marine environmental parameter anomaly forecast results output by the optimized machine learning model, add the denormalized data to the daily climate data, and obtain the marine environmental forecast results for the target area.

2. The marine environment forecasting method according to claim 1, characterized in that: In step 1, the ocean environment parameters include one or more of ocean surface temperature, ocean salinity, sea surface height and ocean current velocity; the spatial position refers to the forecast sea area; and the time range refers to the forecast start time and duration.

3. The marine environment forecasting method according to claim 1, characterized in that: In step 4-1), the neural network model includes one of a convolutional neural network, a long short-term memory neural network and a converter model.

4. The marine environment forecasting method according to claim 1, characterized in that: In step 4-2), the anomalies of ocean environmental parameters are introduced as prediction factors based on dynamic mechanism analysis or existing experience.

5. The marine environment forecasting method according to claim 1, characterized in that: In step 5, after obtaining the marine environment forecast results of the target area, the marine environment forecast results are visualized, including a time series graph and a spatial distribution graph.

Images