A Machine Learning-Based Method and System for Evaluation and Prediction of Marine Three-Dimensional Utilization Resources

Abstract

The application belongs to the technical field of data processing, and specifically discloses a marine three-dimensional utilization resource evaluation and prediction method and system based on machine learning. The method first removes repeated data, processes abnormal values and standardizes multi-dimensional original data of a target sea area to form a standardized data set. Then, multiple sub-regions are divided from the horizontal space, vertical depth and hydrodynamic power dimensions according to sea area characteristic parameters such as the distance from the shore and the water depth. Next, sample data is extracted for each sub-region, and a dedicated machine learning sub-model is constructed and trained. By constructing a contrastive learning loss function, the distribution difference of the sub-model prediction results is calculated, and the parameters are updated through back propagation to realize distribution alignment. Finally, the preprocessed current data is input into the updated sub-model to obtain the first prediction result. The method overcomes the problem of low prediction accuracy caused by the fact that the prior art does not consider the non-homogeneity of the sea area, the sparsity of the sample and the influence of the extreme environment, and can realize accurate resource evaluation and prediction in the whole sea area and in multiple scenarios.

Description

Technical Field

[0001] This invention belongs to the field of data processing technology, specifically relating to a method and system for evaluating and predicting marine three-dimensional utilization resources based on machine learning. Background Technology

[0002] With the rapid development of the global marine economy, the value of its three-dimensional utilization is becoming increasingly prominent, encompassing multiple fields such as aquaculture, offshore wind power, shipping, and ecological protection. Accurate assessment and prediction of marine resources for three-dimensional utilization are a core prerequisite for achieving efficient development and sustainable utilization of marine resources, and are of great significance for promoting high-quality development of the marine economy and ensuring marine ecological security.

[0003] However, marine space exhibits significant heterogeneity, which manifests in multiple dimensions: horizontally, nearshore areas are significantly affected by land-based inputs and human activities, resulting in complex variations in marine environmental parameters, while offshore areas have a relatively stable environment but are difficult to access data for; vertically, different water depths exhibit significant differences in physical and chemical conditions such as light, water pressure, and temperature, directly impacting the distribution of marine life and resource endowment; in terms of hydrodynamics, dynamic processes such as ocean currents and eddies lead to uneven spatial and temporal distribution of marine environmental parameters, further exacerbating the heterogeneity of the sea.

[0004] Existing machine learning-based methods for marine resource assessment and prediction mostly employ a single model and global data training model, failing to fully consider the impact of marine heterogeneity on model prediction performance. On the one hand, in sparsely sampled "marginal areas" (such as offshore ultra-deep water areas and remote eddy zones), the lack of sufficient localized training data prevents the model from accurately learning regional specific characteristics, leading to significant prediction bias. On the other hand, when encountering extreme weather events such as typhoons and El Niño, marine environmental parameters undergo drastic changes, and existing models struggle to adapt quickly to these abrupt shifts. Models trained on historical normal data cannot accurately capture resource change patterns under extreme conditions, resulting in a significant drop in prediction accuracy. Furthermore, traditional methods do not consider the correlation between resource distribution across different regions, leading to independent prediction results for each region. This prevents the formation of a unified and coordinated assessment and prediction system for the entire marine area, making it difficult to meet the decision-making needs of multi-scenario and multi-objective integrated utilization of marine resources.

[0005] Therefore, there is an urgent need for a three-dimensional marine resource utilization assessment and prediction method that can effectively address the problems of marine heterogeneity, sample sparsity, and extreme environmental impacts, so as to improve prediction accuracy and reliability and provide scientific support for marine resource development and utilization. Summary of the Invention

[0006] To address this issue, the present invention provides a method, apparatus, electronic device, medium, and computer program product for evaluating and predicting marine three-dimensional utilization resources based on machine learning, in order to solve the aforementioned technical problems.

[0007] This invention provides a machine learning-based method for evaluating and predicting marine resource utilization across three dimensions, comprising the following steps:

[0008] Based on the multidimensional raw data of the target sea area, the multidimensional raw data is sequentially subjected to duplicate data removal, outlier detection and correction, and data standardization to form a standardized dataset.

[0009] Based on the marine feature parameters in the standardized dataset, the target marine area is divided into multiple sub-regions from three dimensions: horizontal space, vertical depth, and hydrodynamics. The marine feature parameters include distance from shore, water depth, ocean current velocity, and eddy current.

[0010] For each sub-region, sample data of the corresponding sub-region is extracted from the standardized dataset, a machine learning sub-model of the sub-region is constructed, and the machine learning sub-model is trained using the sample data of the corresponding sub-region to obtain a machine learning sub-model for each sub-region after training.

[0011] Construct a contrastive learning loss function, input the prediction results of all trained machine learning sub-models into the contrastive learning loss function, calculate the distribution difference of the prediction results of different sub-models, update the parameters of each sub-model through backpropagation, and align the distribution of the prediction results of machine learning sub-models in each sub-region.

[0012] The current data of the target sea area is collected, preprocessed, and then input into the updated machine learning sub-model of each sub-region to obtain the first prediction result of the three-dimensional utilization resource evaluation of the sea area.

[0013] In another aspect, this application also provides a machine learning-based resource evaluation and prediction system for three-dimensional utilization of marine areas, comprising:

[0014] The target sea area data preprocessing module is used to perform duplicate data removal, outlier detection and correction, and data standardization on the multidimensional raw data of the target sea area to form a standardized dataset.

[0015] The target sea area sub-region division module is used to divide the target sea area into multiple sub-regions based on the sea area feature parameters in the standardized dataset from three dimensions: horizontal space, vertical depth, and hydrodynamics. The sea area feature parameters include distance from the shore, water depth, ocean current velocity, and vorticity.

[0016] The sub-region-specific machine learning sub-model construction and training module is used to extract sample data of the corresponding sub-region from the standardized dataset for each sub-region, construct the machine learning sub-model of the sub-region, and train the machine learning sub-model using the sample data of the corresponding sub-region to obtain the trained sub-region-specific machine learning sub-model.

[0017] The sub-model parameter update module based on the contrastive learning loss function is used to construct the contrastive learning loss function, input the prediction results of all trained machine learning sub-models into the contrastive learning loss function, calculate the distribution difference of the prediction results of different sub-models, update the parameters of each sub-model through backpropagation, and align the distribution of the prediction results of machine learning sub-models in each sub-region.

[0018] The marine three-dimensional utilization resource evaluation and prediction module is used to collect current data of the target marine area, and after preprocessing, input it into the updated machine learning sub-model of each sub-region to obtain the first prediction result of marine three-dimensional utilization resource evaluation.

[0019] This invention constructs dedicated machine learning sub-models for different regions and combines them with a contrastive learning mechanism to align the distribution of prediction results across different areas. This effectively solves the prediction bias problems caused by marine heterogeneity and sample sparsity, significantly improving prediction accuracy in marginal areas such as offshore ultra-deep waters and eddy zones where samples are sparse. Furthermore, by incorporating an adaptive correction mechanism for prediction bias under abnormal environmental conditions, dynamic correction factors are generated using historical extreme event samples and large-scale model inference. This reduces prediction errors under extreme weather conditions such as typhoons and El Niño, avoiding the prediction failure problem of traditional models in extreme environments. Attached Figure Description

[0020] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the embodiments will be briefly introduced below. It should be understood that the following drawings only show some embodiments of the present invention and should not be regarded as a limitation on the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0021] Other features, objects, and advantages of this application will become more apparent from the following detailed description of non-limiting embodiments with reference to the accompanying drawings:

[0022] Figure 1 A flowchart of a machine learning-based method for evaluating and predicting marine three-dimensional utilization resources is provided in an embodiment of the present invention.

[0023] Figure 2 This is a schematic diagram of the CNN-LSTM hybrid neural network structure provided in an embodiment of the present invention.

[0024] Figure 3 This is a schematic diagram of a machine learning-based marine three-dimensional utilization resource evaluation and prediction system provided in an embodiment of the present invention. Detailed Implementation

[0025] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0026] like Figure 1 As shown in the figure, this invention discloses a machine learning-based method for evaluating and predicting marine three-dimensional utilization resources, comprising the following steps:

[0027] Step S1: Based on the multidimensional raw data of the target sea area, the multidimensional raw data is sequentially subjected to duplicate data removal, outlier detection and correction, and data standardization processing to form a standardized dataset. For example, the multidimensional raw data includes sea area geographic location data, marine environmental parameter data, and human activity data.

[0028] Optionally, the marine geographic location data includes, but is not limited to, the marine latitude and longitude coordinates, seabed topography data, and coastline type data; the marine environmental parameter data includes, but is not limited to, seawater temperature, salinity, pH value, dissolved oxygen concentration, chlorophyll concentration, ocean current speed, ocean current direction, wave height, wave period, and tidal level data; the human activity data includes, but is not limited to, fishing operation areas, catch yield, fishing time, offshore wind power platform location, installed capacity, and power generation, shipping routes, vessel types, and number of vessels passing through, and coastal industrial sewage outlet location and sewage discharge data.

[0029] In one embodiment, a multi-source data acquisition system is constructed, consisting of a satellite remote sensing system, an ocean buoy network, survey vessel patrol monitoring, and coastal base monitoring stations, to achieve comprehensive acquisition of multi-dimensional data of the target sea area.

[0030] Specifically, the Gaofen series satellites and Sentinel-3 satellites are used to acquire data on the latitude and longitude coordinates of the sea area, seabed topography (through SAR radar remote sensing technology), coastline type (sandy coast, rocky coast, silty coast), seawater temperature (thermal infrared remote sensing), chlorophyll concentration (ocean color remote sensing), and ocean current speed and direction (Doppler scatterometer remote sensing). For example, the data acquisition frequency is once per day, and the spatial resolution reaches 100m×100m.

[0031] Multiple different types of ocean buoys, including anchored buoys and drifting buoys, are deployed in the target sea area. Anchored buoys are deployed in different areas nearshore, midshore, and offshore. Each buoy is equipped with a temperature sensor, salinity sensor, pH sensor, dissolved oxygen sensor, wave sensor, and tide sensor to collect real-time data on seawater temperature (measurement range -2℃~35℃, accuracy ±0.05℃), salinity (measurement range 0~40psu, accuracy ±0.1psu), pH value (measurement range 6.5~8.5, accuracy ±0.02), dissolved oxygen concentration (measurement range 0~20mg / L, accuracy ±0.1mg / L), wave height (measurement range 0~20m, accuracy ±0.1m), wave period (measurement range 2~30s, accuracy ±0.5s), and tidal level (measurement range -5m~10m, accuracy ±0.01m). For example, the data acquisition frequency is once every 10 minutes.

[0032] A regular survey vessel patrol will be conducted (e.g., monthly) to cover all sub-areas of the target sea area. The survey vessel is equipped with an acoustic Doppler current profiler (ADCP), an underwater robot (ROV), sediment samplers, and biological trawls to collect data on ocean current velocity and direction (vertical resolution 0.5m, horizontal resolution 1km, accuracy ±0.05m / s), seabed sediment types (sand, silt, clay), marine species and abundance (fisheries resource sampling, with one sampling point every 2km), and on-site data on human activities (such as fishing operation records and on-site verification of wind power platform operating parameters).

[0033] Multiple shore-based monitoring stations will be deployed along the coast of the target sea area to focus on monitoring land-based inputs and coastal human activities, including the amount of sewage discharged from coastal industrial outlets (COD and ammonia nitrogen concentration, measurement range 0~500mg / L, accuracy ±5mg / L), the discharge time, the species, area, and yield of aquaculture in nearshore aquaculture areas, and the types and number of ships passing through coastal ports (obtained through a combination of video surveillance and AIS automatic identification system).

[0034] In one embodiment, preprocessing the collected multidimensional raw data to ensure data quality and consistency, exemplarily including:

[0035] Duplicate data removal: A duplicate detection method based on data hash values ​​is used. A unique hash value is calculated for each data entry. If different data entries have the same hash value, they are determined to be duplicate data. The earliest collected data is retained, and the remaining duplicate data is deleted. For example, if the seawater temperature data collected by satellite remote sensing and ocean buoys at the same spatiotemporal point have the same hash value, only the buoy data is retained (because the buoy data has higher accuracy).

[0036] Outlier detection and correction: A method combining the 3σ principle and local weighted regression is adopted. Specifically, for each type of original data, its mean μ and standard deviation σ are calculated, and data that exceed the interval [μ-3σ, μ+3σ] are marked as outliers. For the marked outliers, a local weighted regression algorithm is used, with multiple (e.g., 50) normal data around the outlier as the basis for weight calculation, to fit a corrected value of the outlier and replace the original outlier.

[0037] Data standardization: The Min-Max standardization method is used to map all data to the [0,1] interval. The standardization formula is as follows:

[0038]

[0039] Where x is the original data, It is the minimum value in this type of data. The maximum value in this type of data. This is the standardized data.

[0040] Step S2: Based on the marine feature parameters in the standardized dataset, the target marine area is divided into multiple sub-regions from three dimensions: horizontal space, vertical depth, and hydrodynamics. The marine feature parameters include distance from shore, water depth, ocean current velocity, and eddy current.

[0041] In one embodiment, the division of the horizontal spatial dimension is based on the distance from the shore: the sea area with a distance of 0-30km from the shore is divided into the nearshore area, the sea area with a distance of 30-100km from the shore is divided into the midshore area, and the sea area with a distance greater than 100km from the shore is divided into the offshore area.

[0042] The vertical depth dimension is divided into zones based on water depth: the sea area with a depth of 0-20m is divided into shallow water zone, the sea area with a depth of 20-100m is divided into medium deep water zone, the sea area with a depth of 100-500m is divided into deep water zone, and the sea area with a depth greater than 500m is divided into ultra-deep water zone.

[0043] The regional division of hydrodynamic dimensions is based on ocean current velocity and vorticity: areas with ocean current velocities greater than 1.5 m / s are classified as strong current zones, areas with velocities between 0.5 and 1.5 m / s as medium current zones, and areas with velocities less than 0.5 m / s as weak current zones; areas with absolute vorticity values ​​greater than 10⁻⁻⁶ are classified as weak current zones. 5 The sea area of ​​s⁻¹ is divided into vortex zone, and the rest is non-vortex zone;

[0044] Ultimately, through multi-dimensional superposition, multiple sub-regions were obtained, including the nearshore shallow water area with strong currents, the nearshore shallow water area with medium currents, ..., and the offshore ultra-deep water area with vortex zones.

[0045] Step S3: For each sub-region, extract sample data of the corresponding sub-region from the standardized dataset, construct a machine learning sub-model for the sub-region, and train the machine learning sub-model using the sample data of the corresponding sub-region to obtain a trained machine learning sub-model specific to each sub-region.

[0046] In one embodiment, the machine learning sub-model employs a CNN-LSTM hybrid neural network structure, such as Figure 2 As shown, the CNN-LSTM hybrid neural network includes an input layer, a CNN feature extraction layer, an LSTM temporal processing layer, a fully connected layer, and an output layer.

[0047] The number of neurons in the input layer is consistent with the dimension of the input features, which include standardized marine geographical location features, marine environmental parameter features, and human activity features.

[0048] The CNN feature extraction layer consists of two convolutional layers and two pooling layers. The first convolutional layer has 64 3×3 kernels with ReLU activation function, and the first pooling layer uses 2×2 max pooling. The second convolutional layer has 128 3×3 kernels with ReLU activation function, and the second pooling layer uses 2×2 max pooling.

[0049] The LSTM temporal processing layer consists of two LSTM units, each containing 128 neurons, and uses a Dropout layer to suppress overfitting with a Dropout probability of 0.2. The fully connected layer consists of two layers: the first layer contains 64 neurons with the ReLU activation function, and the second layer contains 32 neurons with the ReLU activation function.

[0050] The output layer contains neurons with the same number as the evaluation and prediction indicators. The evaluation and prediction indicators include fishery resource reserves, suitable aquaculture area, wind power development potential index, and shipping capacity. The activation function is a linear activation function.

[0051] In one embodiment, the training process of the machine learning sub-model includes:

[0052] The sample data of the corresponding sub-regions were divided into training and validation sets in a 7:3 ratio. The training batch size was set to 64, the maximum number of training iterations was set to 5000, and the Adam optimization algorithm was used. The initial learning rate was 0.001, the momentum parameter β1=0.9, the second-order momentum parameter β2=0.999, and the numerical stability parameter ε=1e-8. The mean squared error (MSE) between the predicted and actual values ​​was used as the loss function. During training, the MSE of the validation set was calculated every 100 iterations. If the MSE of the validation set did not decrease for 200 consecutive iterations, training was stopped, and the current model parameters were saved as the completed machine learning sub-model.

[0053] Step S4: Construct a contrastive learning loss function, input the prediction results of all trained machine learning sub-models into the contrastive learning loss function, calculate the distribution difference of the prediction results of different sub-models, update the parameters of each sub-model through backpropagation, and align the distribution of the prediction results of machine learning sub-models in each sub-region.

[0054] In one embodiment, the contrastive learning loss function employs a hybrid loss function combining KL divergence and contrastive loss, expressed as:

[0055]

[0056] in, Here, is the weighting coefficient, exemplified by a value of 0.6; KL(P||Q) is the KL divergence between the prediction distributions P and Q of the two sub-models; and N is the number of sub-models. As labels, when the sub-region types corresponding to sub-model i and sub-model j are similar, ,otherwise , is the Euclidean distance between the prediction results of sub-model i and sub-model j; D is the marginal parameter, which is exemplarily set to 2.0;

[0057] It is understandable that this is being done when calculating KL divergence and Euclidean distance. Previously, the prediction results could be normalized (e.g., Min-Max standardization, Z-Score standardization) to eliminate dimensional differences, which will not be elaborated here.

[0058] In one embodiment, by minimizing the The parameters of the fully connected layers and LSTM layers of each sub-model are updated to make the predicted distributions of sub-models in similar sub-regions more similar, and to maintain a reasonable degree of differentiation in the predicted distributions of sub-models in different sub-regions.

[0059] In one embodiment, for example, the similarity of sub-regions is calculated from two dimensions: environmental characteristics (water temperature, salinity, water depth, ocean currents, etc.) and resource types (aquaculture, wind power development, shipping, etc.). Environmental characteristic similarity is obtained by calculating the cosine similarity of the mean vectors of environmental parameters between the two sub-regions. Environmental parameters include seawater temperature, salinity, ocean current velocity, and water depth. Resource type similarity is determined based on the main resource types of the sub-regions (such as aquaculture, wind power development, shipping, etc.). If the main resource types of the two sub-regions are the same, the resource type similarity is 1; otherwise, it is 0.

[0060] Specifically, for example, for each sub-region, the seawater temperature is extracted. ,salinity Ocean current speed , water depth The mean vector of environmental parameters, for example, the mean vector of sub-region i is:

[0061]

[0062] The cosine similarity is used to calculate the environmental feature similarity between sub-regions i and j. :

[0063]

[0064] in, For vector dot product, The vector magnitude is [0, 1], and the closer the value is to 1, the more similar the environmental features are.

[0065] For resource type similarity Based on the core resource types for the three-dimensional utilization of marine areas (such as aquaculture, wind power development, navigation, and ecological protection), the primary resource type is labeled for each sub-region (e.g., sub-region i is labeled as "Aquaculture Dominant Area"). If the primary resource types of sub-regions i and j are the same, ,otherwise, .

[0066] By weighted fusion of environmental characteristics and resource type similarity, the comprehensive similarity scores of sub-regions i and j are obtained:

[0067]

[0068] in, For example, the weighting coefficients (which can be adjusted according to the characteristics of the sea area) are as follows: For example, when At that time, the judgment (Similar sub-regions), otherwise (Different sub-regions).

[0069] For calculating the KL divergence of the prediction results, for example, corresponding to the sub-model (CNN-LSTM sub-model), the output prediction result vector is obtained to obtain the prediction distribution of each sub-model. ;

[0070] Specifically, for the result vector output by the sub-model, the discrete vector values ​​are transformed into a continuous distribution of prediction results through probability density estimation. Specifically, for example, it includes:

[0071] The set of result vectors output by the same sub-model under multiple training / test samples Where m is the sample size, standardization (e.g., Z-score standardization) is performed to eliminate the influence of differences in indicator dimensions on the distribution estimation. A kernel density estimation algorithm (KDE) is used, employing the preprocessed result vector set as samples to construct a continuous probability density function (i.e., the predicted result distribution). ).

[0072] The kernel function is chosen from Gaussian kernels (to satisfy the smoothness assumption of marine evaluation indicators), and the bandwidth is optimized through cross-validation (such as 5-fold cross-validation) to ensure the accuracy of the distribution estimation; the resulting probability density function... The distribution of prediction results for this sub-model is represented by x, which is the value of the evaluation index.

[0073] Next, calculate for each pair of similar sub-regions The formula for measuring the difference between the predicted distributions of the two distributions (the smaller the value, the more similar the distributions) is:

[0074]

[0075] in, , Predict the probability of the sub-model being distributed at sample point x.

[0076] In one embodiment, for backpropagation, it is, for example, to To optimize the objective, the parameters of the fully connected and LSTM layers in the sub-model are updated (the parameters of the CNN layers are not updated, as they are responsible for extracting general spatial features and need to remain stable to avoid disrupting their ability to extract these features; the fully connected and LSTM layers are responsible for learning sub-region-specific patterns and need to be aligned). Gradient descent is used to adjust the parameters to minimize... For similar sub-regions, reduce KL divergence and lower [the threshold for improvement]. For the different sub-regions, ensure To maintain differentiation.

[0077] When the value on the validation set no longer decreases, or the alignment of the predicted distributions of similar sub-regions (e.g., KL divergence mean (<0.1), the discriminative power of dissimilar sub-regions () Training should be stopped when the percentage of participants is greater than 95% (meeting the requirement).

[0078] Therefore, by comparing and learning the loss function, after aligning the predicted distributions of similar sub-regions, the model can transfer the patterns of densely sampled sub-regions to sparsely sampled sub-regions (e.g., in two similar sub-regions, one has dense samples and the other has sparse samples). Simultaneously, the discriminative constraint on dissimilar sub-regions prevents the model from confusing the characteristics of different regions. This allows the model to learn common patterns across regions while respecting the differences between sub-regions, ultimately improving the accuracy and robustness of resource assessment and prediction across the entire sea area.

[0079] Step S5: Collect current data of the target sea area, preprocess it, and input it into the updated machine learning sub-model of each sub-region to obtain the first prediction result of the three-dimensional utilization resource evaluation of the sea area.

[0080] In one embodiment, target sea area data is collected and the collected data is preprocessed in the same manner as in the foregoing embodiments, which will not be repeated here.

[0081] Subsequently, by using the spatial coordinates (latitude and longitude, water depth) of real-time data, the corresponding sub-region code is matched (such as nearshore aquaculture area, offshore shipping area, etc., refer to the regional division rules above); the machine learning sub-model specific to that sub-region is called (such as a CNN-LSTM hybrid network, a model optimized through comparative learning). Based on the region where the sub-model is located, the corresponding resource evaluation result is output.

[0082] Optionally, an extreme event sample database is constructed based on target sea area data during historical extreme climate events. A pre-trained large language model is used to extract features and infer meaning from the data in the extreme event sample database. Combined with the actual prediction deviation under extreme events, a correction factor calculation model is trained. The pre-trained large language model is a marine-specific model based on the Transformer architecture, obtained through pre-training on massive amounts of marine observation data and marine literature corpora. The correction factor calculation model uses a gradient boosting decision tree algorithm, taking the feature vector output by the large language model as input and the actual prediction deviation as the output label to train the correction factor calculation model.

[0083] In this embodiment, historical extreme climate event data is collected through a multi-source data acquisition system, covering event types such as typhoons, El Niño, La Niña, severe cold waves, torrential rains, storm surges, and red tides. For example, data sources include:

[0084] Ocean observation network: Historical monitoring data from multiple ocean buoys (including anchored buoys and drifting buoys), underwater sensor arrays, and shore-based radars within the target sea area, acquiring marine environmental parameters such as seawater temperature, salinity, pH value, dissolved oxygen concentration, ocean current speed / direction, wave height / period, and tidal level for 3 months before the event, during the event, and 1 month after the event, with a temporal resolution of 10 minutes / time and a spatial resolution of 100m×100m.

[0085] Fisheries and marine development records: historical data from coastal fishing port management systems, offshore wind power operation platforms, and shipping management AIS systems, including data on human activities and resource utilization during the event period, such as fishery catch / operation area, wind power generation / equipment status, number of vessels passing through / route adjustments.

[0086] Meteorological and marine forecast archives: historical data from the National Meteorological Center and marine forecasting stations, supplemented with information such as event type (e.g., typhoon-strong typhoon), intensity (typhoon wind speed, El Niño sea surface temperature anomaly), and affected area (latitude and longitude coverage area).

[0087] Next, duplicate and invalid data (such as outliers caused by sensor malfunctions) are removed, and outliers are corrected by using the 3σ principle combined with local weighted regression (such as the sudden change in ocean current velocity data of a buoy during a typhoon, which is corrected by interpolation of surrounding normal data).

[0088] For each sample, a standardized labeling system is established, including the event type (e.g., "Typhoon - Severe Typhoon"), intensity level (typhoons are classified into tropical storms to super typhoons based on wind speed), impact range (latitude and longitude boundaries), and resource impact degree (fishery resource loss rate, aquaculture facility damage rate). For example, a typhoon sample is labeled as: "Event type: Typhoon - Severe Typhoon; Intensity: Wind speed 45 m / s; Impact range: 30-32°N, 121-123°E; Resource impact degree: Fishery resource loss rate 25%, aquaculture facility damage rate 30%."

[0089] In this embodiment, the large language model based on the Transformer architecture is initialized with parameters based on a general language model (such as BERT-Base) and pre-trained using the following marine domain data:

[0090] Massive ocean observation data: Argo buoy global ocean profile dataset (including temperature and salinity profile data, over 1 million records) and CMIP6 ocean model simulation dataset (covering ocean environmental parameters under different climate scenarios, over 500,000 records), providing the spatiotemporal distribution patterns of ocean environmental parameters.

[0091] Marine literature corpus: Oceanographic papers and marine industry reports indexed by Web of Science, CNKI, etc., covering professional knowledge and semantic logic in fields such as marine ecology, resource development, and disaster early warning.

[0092] Next, the model is trained, which, exemplarily, includes pre-training task design and training strategies. Specifically,

[0093] For example, the design of pre-training tasks includes:

[0094] Masked Language Modeling (MLM): Randomly masking marine environmental parameters and technical terms (such as "thermocline" and "upwell") allows the model to learn semantic associations in the marine domain and predict the correct vocabulary for masking locations.

[0095] Sequence prediction task: Input time series of marine environmental parameters, let the model predict the parameter changes in the next period, and learn the temporal patterns of marine data.

[0096] For training strategies, for example:

[0097] A freeze-thaw strategy is adopted: the parameters of the bottom 6 layers of Transformer are frozen in the first 1000 iterations, and only the top 6 layers are trained with a learning rate of 0.0001; in the next 1000 iterations, all layers are unfrozen, the whole model is fine-tuned, the batch size is 32, and the optimizer is AdamW.

[0098] In this embodiment, data from the extreme event sample library is input into a pre-trained large language model, outputting a 1024-dimensional feature vector that covers the environmental features and semantic relationships of the event (such as the potential relationship between sea temperature, ocean currents, and fishery losses during typhoons). Principal component analysis (PCA) is used to reduce the dimensionality of the feature vector, retaining 95% of the information, resulting in a 64-dimensional feature vector. Then, Min-Max normalization is applied to map it to the [0,1] interval to eliminate the influence of dimensions, which is then used as input to the correction factor calculation model.

[0099] Next, the actual prediction bias under extreme events is calculated: the data during the event is input into a sub-regional machine learning model (such as a CNN-LSTM sub-model) to obtain preliminary prediction values ​​(such as fishery resource reserves); these are compared with the actual resource assessment results after the event (such as actual catch), and the difference is the actual prediction bias (label value). For example, if the preliminary prediction of fishery resource reserves is 500 tons and the actual value is 400 tons, then the bias is 100 tons.

[0100] Next, using the feature vector output by the large language model as input and the actual prediction bias as the output label, the gradient boosting decision tree (GBDT) model is trained to obtain the correction factor calculation model.

[0101] Specifically, for example, the GBDT model has 200 decision trees, a maximum depth of 8 per tree, a minimum number of sample splits of 20, a minimum number of leaf nodes per sample of 10, a learning rate of 0.01, and a loss function of mean squared error (MSE). The parameters are optimized through grid search, with the goal of minimizing the MSE of the validation set (sample data split in a 7:3 ratio), to determine the optimal parameter combination.

[0102] Optionally, based on the current data of the target sea area collected in real time, it is determined whether the current environment is abnormal. If so, the correction factor calculation model is called to generate a correction factor, and the final sea area three-dimensional utilization resource evaluation prediction result is output after correcting the first prediction result.

[0103] In one embodiment, the wind speed, ocean current speed, and sea temperature of the sea area are monitored in real time. If the wind speed is greater than level 12 (typhoon level), the sudden change in ocean current speed exceeds 50%, the weekly change in sea temperature exceeds 3°C, or an extreme weather warning signal is received from the meteorological department, it is determined that the current environment is under abnormal conditions.

[0104] When the target sea area is determined to be under abnormal environmental conditions, the following correction procedure is executed, for example:

[0105] Feature vector generation: Real-time data under abnormal environments (including environmental parameters and event annotation information) are input into a Transformer model pre-trained with massive amounts of marine data, outputting a 1024-dimensional feature vector. After PCA dimensionality reduction (retaining 95% of the information) and Min-Max standardization, a 64-dimensional input feature vector is obtained. .

[0106] Corrected factor prediction: Input a pre-trained GBDT correction factor calculation model, which is based on the pre-trained extreme event feature-prediction bias mapping relationship, and output a correction factor. .

[0107] After obtaining the first prediction result from the machine learning sub-model of the sub-region, the following formula is used for correction:

[0108]

[0109] in, The initial prediction results output by the sub-model. This is the final revised forecast result. If the revision factor is negative, it means that the initial forecast overestimated the actual value (e.g., typhoons caused losses to fishery resources, so the forecast needs to be revised downward); if it is positive, it means that the initial forecast underestimated the actual value (e.g., El Niño caused enrichment of certain fishery resources, so the forecast needs to be revised upward).

[0110] Therefore, by using a pre-trained large language model to mine event features and combining it with the GBDT correction factor model, the preliminary prediction results are dynamically adjusted to solve the prediction bias problem of traditional models in extreme environments and adapt to the complex needs of marine three-dimensional utilization resource evaluation.

[0111] Figure 3 A machine learning-based resource evaluation and prediction system 300 for three-dimensional utilization of marine areas is shown. This device embodiment is similar to... Figure 1 Corresponding to the illustrated method embodiments, the specific methods include:

[0112] The target sea area data preprocessing module 301 is used to perform duplicate data removal, outlier detection and correction, and data standardization processing on the multidimensional raw data of the target sea area to form a standardized dataset.

[0113] The target sea area sub-region division module 302 is used to divide the target sea area into multiple sub-regions based on the sea area feature parameters in the standardized dataset from three dimensions: horizontal space, vertical depth, and hydrodynamics. The sea area feature parameters include distance from the shore, water depth, ocean current velocity, and vorticity.

[0114] The sub-region-specific machine learning sub-model construction and training module 303 is used to extract sample data of the corresponding sub-region from the standardized dataset for each sub-region, construct the machine learning sub-model of the sub-region, and train the machine learning sub-model using the sample data of the corresponding sub-region to obtain the trained sub-region-specific machine learning sub-model.

[0115] The sub-model parameter update module 304 based on the contrastive learning loss function is used to construct the contrastive learning loss function, input the prediction results of all trained machine learning sub-models into the contrastive learning loss function, calculate the distribution difference of the prediction results of different sub-models, update the parameters of each sub-model through backpropagation, and align the distribution of the prediction results of machine learning sub-models in each sub-region.

[0116] The marine three-dimensional utilization resource evaluation and prediction module 305 is used to collect the current data of the target marine area, and after preprocessing, input it into the updated machine learning sub-model of each sub-region to obtain the first prediction result of marine three-dimensional utilization resource evaluation.

[0117] In the above embodiments, the descriptions of each embodiment have different focuses. For parts that are not described in detail or recorded in a certain embodiment, please refer to the relevant descriptions of other embodiments.

[0118] The above-described embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application, and should all be included within the protection scope of this application.

Claims

1. A method for predicting the evaluation of marine stereoscopic utilization resources based on machine learning, characterized in that, The method comprises the following steps: Based on the multi-dimensional original data of the target sea area, the multi-dimensional original data is sequentially subjected to repeated data elimination, abnormal value detection and correction, and data standardization processing to form a standardized data set; Based on the sea area characteristic parameters in the standardized data set, the target sea area is regionally divided from three dimensions of horizontal space, vertical depth and hydrodynamic power to obtain a plurality of sub-regions, and the sea area characteristic parameters include offshore distance, water depth, sea current velocity and vorticity; For each sub-region, sample data of the corresponding sub-region is extracted from the standardized data set, a machine learning sub-model of the sub-region is constructed, the machine learning sub-model is trained using the sample data of the corresponding sub-region, and a trained exclusive machine learning sub-model of each sub-region is obtained; A contrast learning loss function is constructed, the prediction results of all trained machine learning sub-models are input into the contrast learning loss function, the distribution difference of the prediction results of different sub-models is calculated, each sub-model parameter is updated through back propagation, and the distribution of the prediction results of each sub-region machine learning sub-model is aligned; Current data of the target sea area is collected, preprocessed and input into the updated machine learning sub-model of each sub-region to obtain a sea area three-dimensional utilization resource evaluation first prediction result; The contrast learning loss function adopts a hybrid loss function based on the combination of KL divergence and contrast loss, and the expression is: wherein, is a weight coefficient, KL(P||Q) is the KL divergence of the prediction result distribution P and Q of two sub-models, N is the number of sub-models, is a label, when the sub-region types corresponding to the sub-model i and the sub-model j are similar, , otherwise , is the Euclidean distance of the prediction results of the sub-model i and the sub-model j, and D is a marginal parameter. 2.The method of claim 1, wherein, Further comprising, Based on the target sea area data during the historical extreme climate event, an extreme event sample library is constructed; a pre-trained large language model is used to extract features and reason the data in the extreme event sample library, and a correction factor calculation model is trained in combination with the actual prediction bias under the extreme event.

3. The machine learning-based sea area three-dimensional utilization resource evaluation prediction method according to claim 2, characterized in that, Based on the real-time collected current data of the target sea area, it is judged whether the current is in an abnormal environmental condition, if yes, the correction factor calculation model is called to generate a correction factor, and the first prediction result is corrected to output a final sea area three-dimensional utilization resource evaluation prediction result. 4.The method of claim 3, wherein, Comprise: The machine learning sub-model adopts a CNN-LSTM hybrid neural network structure, the CNN-LSTM hybrid neural network comprises an input layer, a CNN feature extraction layer, an LSTM time sequence processing layer, a full connection layer and an output layer; the number of neurons of the input layer is consistent with the input feature dimension, the CNN feature extraction layer comprises two convolution layers and two pooling layers, the LSTM time sequence processing layer is provided with two LSTM units, and the output layer comprises neurons consistent with the number of evaluation prediction indexes.

5. The machine learning-based sea area three-dimensional utilization resource evaluation prediction method according to claim 4, characterized in that, The input features include standardized sea area geographical location features, marine environment parameter features and human activity features; the evaluation prediction indexes include fishery resource reserves, suitable aquaculture area, wind power development potential index and shipping capacity, and the activation function adopts a linear activation function.

6. The method according to claim 1, wherein the method is characterized in that, by minimizing the , update the parameters of the full connection layer and the LSTM layer of each sub-model, so that the prediction distribution of the sub-models of similar sub-regions is closer, and the prediction distribution of the sub-models of different sub-regions remains reasonable discrimination.

7. The method according to claim 6, wherein the method is characterized in that, The pre-trained large language model is a marine field special model based on the Transformer architecture, which is obtained by pre-training on marine observation data and marine literature corpus.

8. The method according to claim 7, wherein the method is characterized in that, The correction factor calculation model adopts the gradient boosting decision tree algorithm, takes the feature vector output by the large language model as input, and takes the actual prediction deviation as output label to train the correction factor calculation model. 9.A machine learning based system for predicting the evaluation of marine stereoscopic utilization resources, characterized in that, The method comprises: A target sea area data preprocessing module is configured to perform repeated data elimination, abnormal value detection and correction, and data standardization processing on multi-dimensional original data of a target sea area to form a standardized data set; A target sea area sub-region division module is configured to divide the target sea area into a plurality of sub-regions from horizontal space, vertical depth, and hydrodynamic dimensions based on sea area characteristic parameters in the standardized data set, wherein the sea area characteristic parameters include offshore distance, water depth, sea current velocity, and vorticity; A sub-region exclusive machine learning sub-model construction and training module is configured to extract sample data of each sub-region from the standardized data set, construct a machine learning sub-model for the sub-region, and train the machine learning sub-model using the sample data of the corresponding sub-region to obtain trained sub-region exclusive machine learning sub-models; A sub-model parameter updating module based on a contrastive learning loss function is configured to construct a contrastive learning loss function, input prediction results of all trained machine learning sub-models into the contrastive learning loss function, calculate distribution differences of prediction results of different sub-models, update parameters of each sub-model by back propagation, and align distributions of prediction results of each sub-region machine learning sub-model; A sea area three-dimensional utilization resource evaluation and prediction module is configured to collect current data of a target sea area, input the preprocessed data into updated machine learning sub-models of each sub-region, and obtain a first prediction result of sea area three-dimensional utilization resource evaluation; The contrastive learning loss function adopts a hybrid loss function based on the combination of KL divergence and contrastive loss, and the expression is as follows: wherein, is a weight coefficient, KL(P||Q) is the KL divergence of the prediction result distribution P and Q of two sub-models, N is the number of sub-models, is a label, when the sub-region types corresponding to the sub-model i and the sub-model j are similar, , otherwise , is the Euclidean distance of the prediction results of the sub-model i and the sub-model j, and D is a marginal parameter.

Images