Sea area three-dimensional utilization resource evaluation and prediction method and system based on machine learning

Abstract

The invention belongs to the technical field of data processing, and particularly discloses a sea area three-dimensional utilization resource evaluation prediction method and system based on machine learning, and the method comprises the steps: firstly carrying out the repeated data elimination, abnormal value processing and standardization of multi-dimensional original data of a target sea area, and forming a standardized data set; dividing a plurality of sub-regions from three dimensions of horizontal space, vertical depth and hydrological power according to sea area characteristic parameters such as offshore distance and water depth; extracting sample data for each sub-region, and constructing and training an exclusive machine learning sub-model; calculating the distribution difference of prediction results of the sub-models by constructing a contrast learning loss function, and performing back propagation on updated parameters to realize distribution alignment; and finally, collecting current data, preprocessing, and inputting the current data into the updated sub-model to obtain a first prediction result. According to the method, the problems of low prediction precision and the like caused by the fact that sea area heterogeneity, sample sparseness and extreme environment influence are not considered in the prior art are solved, and accurate resource evaluation prediction in the whole sea area and multiple scenes can be realized.

Description

TECHNICAL FIELD

[0001] The present application belongs to the technical field of data processing, and particularly relates to a marine three-dimensional utilization resource evaluation and prediction method and system based on machine learning. BACKGROUND

[0002] With the rapid development of global marine economy, its three-dimensional utilization value is increasingly prominent, covering multiple fields such as fish farming, offshore wind power, shipping transportation, and ecological protection. Accurate marine three-dimensional utilization resource evaluation and prediction is the core prerequisite for efficient development and sustainable utilization of marine resources, and is of great significance to promoting high-quality development of marine economy and ensuring marine ecological safety. However, marine space has significant heterogeneous characteristics, which are reflected in multiple dimensions. In the horizontal direction, the nearshore area is significantly affected by land source input and human activities, and the marine environment parameters change complexly, while the offshore area has a relatively stable environment but it is difficult to obtain data. In the vertical direction, the physical and chemical conditions such as light, water pressure, and temperature are obviously different in different water depth areas, which directly affect the distribution of marine organisms and resource endowment. In terms of hydrodynamic power, the dynamic processes such as ocean currents and eddies lead to uneven spatiotemporal distribution of marine environmental parameters, further exacerbating the heterogeneity of the marine space.

[0003] Existing marine resource evaluation and prediction methods based on machine learning mostly use single models and global data training mode, without fully considering the influence of marine heterogeneity on model prediction performance. On the one hand, in the "edge area" (such as offshore ultra-deep water area and remote eddy area) with sparse samples, the model cannot accurately learn the regional specificity due to the lack of sufficient localized training data, resulting in large prediction errors. On the other hand, when encountering extreme climate events such as typhoons and El Nino, the marine environmental parameters will change dramatically, and the existing model is difficult to quickly adapt to such mutations. The model trained based on historical normal data cannot accurately capture the resource change rule under extreme conditions, and the prediction accuracy is greatly reduced. In addition, the traditional method does not consider the correlation of resource distribution in different regions, and the prediction results of each region are independent of each other, which cannot form a unified and coordinated evaluation and prediction system for the whole marine space, and it is difficult to meet the decision-making needs of multi-scenario and multi-target of marine three-dimensional utilization.

[0004] Therefore, there is an urgent need for a marine three-dimensional utilization resource evaluation and prediction method that can effectively deal with the problems of marine heterogeneity, sample sparsity, and extreme environmental impact, in order to improve the prediction accuracy and reliability, and provide scientific support for marine resource development and utilization. SUMMARY

[0005] To this end, the present application provides a marine three-dimensional utilization resource evaluation and prediction method and device based on machine learning, an electronic device, a medium, and a computer program product, to solve the above technical problems.

[0006] The application provides a sea area three-dimensional utilization resource evaluation prediction method based on machine learning, comprising the following method 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 flow rate 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 sample data of the corresponding sub-region is used to train the machine learning sub-model, 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, the parameters of each sub-model are 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.

[0007] Another aspect of the application also provides a sea area three-dimensional utilization resource evaluation prediction system based on machine learning, comprising: A target sea area data preprocessing module is configured to, based on multi-dimensional original data of a target sea area, sequentially perform repeated data elimination, abnormal value detection and correction, and data standardization processing on the multi-dimensional original data to form a standardized data set; A target sea area sub-region division module is configured to, based on sea area characteristic parameters in the standardized data set, regionally divide the target sea area 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 flow rate and vorticity; A sub-region exclusive machine learning sub-model construction and training module is configured to, for each sub-region, extract sample data of the corresponding sub-region from the standardized data set, construct a machine learning sub-model of the sub-region, train the machine learning sub-model using the sample data of the corresponding sub-region, and obtain a trained exclusive machine learning sub-model of each sub-region; The sub-model parameter updating module based on the contrast learning loss function is used for constructing a contrast learning loss function, inputting the prediction results of all trained machine learning sub-models into the contrast learning loss function, calculating the distribution difference of the prediction results of different sub-models, updating the parameters of each sub-model through back propagation, and aligning the distribution of the prediction results of each sub-regional machine learning sub-model. The sea area three-dimensional utilization resource evaluation prediction module is used for collecting current data of a target sea area, inputting the preprocessed data into the updated machine learning sub-models of each sub-region, and obtaining a first prediction result of sea area three-dimensional utilization resource evaluation.

[0008] The application constructs exclusive machine learning sub-models by sub-regions, aligns the distribution of prediction results in different regions by combining a contrast learning mechanism, effectively solves the prediction deviation problem caused by the non-homogeneity of sea areas and sample sparsity, and obviously improves the prediction accuracy in edge regions such as offshore ultra-deep water areas and vortex areas with sample sparsity. Meanwhile, by combining a prediction deviation self-adaptive correction mechanism under abnormal environmental conditions, using historical extreme event samples and a large model to generate a dynamic correction factor, the prediction error under extreme climate such as typhoon and El Nino is reduced, and the prediction failure problem of traditional models under extreme environment is avoided. BRIEF DESCRIPTION OF DRAWINGS

[0009] In order to more clearly illustrate the technical solutions of the embodiments of the application, the following will briefly introduce the drawings needed to be used in the embodiments. It should be understood that the following drawings only show some embodiments of the application, and therefore should not be regarded as a limitation to the scope. For those skilled in the art, other related drawings can also be obtained without creative labor on the basis of these drawings.

[0010] Other features, objects and advantages of the application will become more apparent from the following detailed description of non-limiting embodiments, made with reference to the following drawings: Figure 1 A flow chart of a sea area three-dimensional utilization resource evaluation prediction method based on machine learning is provided for the embodiments of the application.

[0011] Figure 2 A CNN-LSTM hybrid neural network structure schematic diagram is provided for the embodiments of the application.

[0012] Figure 3 A sea area three-dimensional utilization resource evaluation prediction system schematic diagram based on machine learning is provided for the embodiments of the application. DETAILED DESCRIPTION

[0013] To make the purposes, technical solutions, and advantages of the embodiments of the present application clearer, the technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are some but not all of the embodiments of the present application. Based on the embodiments in the present application, all other embodiments obtained by a person of ordinary skill in the art without creative work fall within the protection scope of the present application.

[0014] As shown in Figure 1 The embodiments of the present application disclose a marine three-dimensional utilization resource evaluation and prediction method 100 based on machine learning, comprising the following method steps: Step S1, 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. Exemplarily, the multi-dimensional original data includes sea area geographic location data, marine environment parameter data, and human activity data.

[0015] Optionally, the sea area geographic location data includes but is not limited to sea area latitude and longitude coordinates, seabed topography data, and coastline type data; the marine environment parameter data includes but is not limited to seawater temperature, salinity, pH value, dissolved oxygen concentration, chlorophyll concentration, sea current speed, sea current direction, wave height, wave period, and tidal water level data; and the human activity data includes but is not limited to fishing operation area, fishing yield, fishing time, offshore wind power platform location, installed capacity, power generation, shipping route, ship type, ship passage number, coastal industrial sewage outlet location, and sewage discharge data.

[0016] In one embodiment, a multi-source data acquisition system composed of satellite remote sensing systems, marine buoy networks, survey ship cruising monitoring, and coastal shore-based monitoring stations is constructed to realize comprehensive acquisition of multi-dimensional data of the target sea area.

[0017] Specifically, high-resolution series satellites, Sentinel-3 satellites, etc. are used to acquire latitude and longitude coordinates of the sea area, seabed topography (through SAR radar remote sensing technology), coastline types (sand coast, rock coast, and silt coast), seawater temperature (thermal infrared remote sensing), chlorophyll concentration (ocean color remote sensing), and sea current speed and direction (Doppler scatterometer remote sensing) data. Exemplarily, the data acquisition frequency is 1 time / day, and the spatial resolution reaches 100 m x 100 m.

[0018] In the target sea area, multiple different types of ocean buoys are deployed, including anchored buoys and drifting buoys. Anchored buoys are deployed in different areas near the shore, in the middle of the shore, and far from the shore. Each buoy is equipped with a temperature sensor, a salinity sensor, a pH sensor, a dissolved oxygen sensor, a wave sensor, and a tide sensor to collect real-time data on seawater temperature (measuring range -2℃~35℃, accuracy ±0.05℃), salinity (measuring range 0~40psu, accuracy ±0.1psu), pH value (measuring range 6.5~8.5, accuracy ±0.02), dissolved oxygen concentration (measuring range 0~20mg / L, accuracy ±0.1mg / L), wave height (measuring range 0~20m, accuracy ±0.1m), wave period (measuring range 2~30s, accuracy ±0.5s), and tidal level (measuring range -5m~10m, accuracy ±0.01m). The data collection frequency is 1 time / 10 minutes.

[0019] Regularly conduct a survey ship cruise (e.g., every month) covering all sub-areas of the target sea area. The survey ship is equipped with an acoustic Doppler current profiler (ADCP), an underwater robot (ROV), a sediment sampler, a biological trawl net, and other equipment to collect data on sea current speed and direction (vertical resolution 0.5m, horizontal resolution 1km, accuracy ±0.05m / s), seabed sediment type (sand, silt, clay), marine species and quantity (fishery resource sampling, 1 sampling point every 2km), and human activity site data (such as fishery catch operation records, wind power platform operation parameter on-site verification).

[0020] Along the coast of the target sea area, multiple shore-based monitoring stations are deployed to monitor land source input and coastal human activity data, including the amount of pollution (COD, ammonia nitrogen concentration, measuring range 0~500mg / L, accuracy ±5mg / L) and time of pollution from coastal industrial discharge outlets, the species, area, and yield of nearshore aquaculture areas, and the type and number of ships passing through coastal ports (obtained through video monitoring combined with the AIS ship automatic identification system).

[0021] In one embodiment, the collected multi-dimensional raw data is preprocessed to ensure data quality and consistency. Specifically, it includes: Duplicate data elimination: A duplicate detection method based on data hash value is used to calculate a unique hash value for each data entry. If the hash values of different data entries are the same, it is determined that the data is duplicated, and the earliest collected data is retained and the rest of the duplicated data is deleted. For example, if the satellite remote sensing and the ocean buoy collect seawater temperature data at the same spatiotemporal point, and the hash values are consistent, only the buoy data is retained (because the buoy data has higher accuracy).

[0022] Outlier detection and correction: a method based on 3σ principle and local weighted regression is adopted, specifically: for each type of original data, calculate its mean μ and standard deviation σ, mark the data outside the interval [μ-3σ, μ+3σ] as outliers; for the marked outliers, use the local weighted regression algorithm to calculate the correction value of the outliers based on the weight of multiple (e.g. 50) normal data around the outliers, and replace the original outliers with the correction value; Data standardization: Min-Max standardization method is adopted to map all data to the interval [0, 1], and the standardization formula is:

[0023] where x is the original data, is the minimum value of the type of data, is the maximum value of the type of data, is the standardized data.

[0024] Step S2, based on the sea area characteristic parameters in the standardized data set, the target sea area is regionally divided from horizontal space, vertical depth, and hydrodynamic power three dimensions to obtain multiple sub-regions, and the sea area characteristic parameters include offshore distance, water depth, sea current velocity, and vorticity.

[0025] In one embodiment, the horizontal space dimension regional division is based on offshore distance: the sea area with offshore distance of 0-30 km is divided into nearshore region, the sea area with offshore distance of 30-100 km is divided into middle shore region, and the sea area with offshore distance greater than 100 km is divided into offshore region; The vertical depth dimension regional division is based on water depth: the sea area with water depth of 0-20 m is divided into shallow water area, the sea area with water depth of 20-100 m is divided into middle deep water area, the sea area with water depth of 100-500 m is divided into deep water area, and the sea area with water depth greater than 500 m is divided into super deep water area; The hydrodynamic power dimension regional division is based on sea current velocity and vorticity: the sea area with sea current velocity greater than 1.5 m / s is divided into strong current area, the sea area with sea current velocity of 0.5-1.5 m / s is divided into middle current area, and the sea area with sea current velocity less than 0.5 m / s is divided into weak current area; the sea area with absolute value of vorticity greater than 10⁻ 5 s⁻¹ is divided into vortex area, and the rest is non-vortex area; Finally, through multi-dimensional superposition, nearshore shallow water area strong current area, nearshore shallow water area middle current area, …, offshore super deep water area vortex area, and multiple sub-regions are obtained.

[0026] Step S3, for each sub-region, sample data corresponding to the 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 sample data corresponding to the sub-region, and a trained sub-region exclusive machine learning sub-model is obtained.

[0027] In one embodiment, the machine learning sub-model adopts a CNN-LSTM hybrid neural network structure, as shown in Figure 2 The CNN-LSTM hybrid neural network includes an input layer, a CNN feature extraction layer, an LSTM time series processing layer, a full connection layer, and an output layer, as shown. The number of neurons of the input layer is consistent with the input feature dimension, and the input features include standardized sea area geographical location features, marine environment parameter features, and human activity features. The CNN feature extraction layer includes two convolution layers and two pooling layers. The first convolution layer is provided with 64 3x3 size convolution kernels, and the activation function is ReLU. The first pooling layer adopts 2x2 maximum pooling. The second convolution layer is provided with 128 3x3 size convolution kernels, and the activation function is ReLU. The second pooling layer adopts 2x2 maximum pooling. The LSTM time series processing layer is provided with two layers of LSTM units, each layer containing 128 neurons, and a Dropout layer is used to suppress overfitting with a Dropout probability of 0.2. The full connection layer is provided with two layers, the first layer containing 64 neurons with a ReLU activation function, and the second layer containing 32 neurons with a ReLU activation function. The output layer contains neurons consistent with the number of evaluation prediction indicators, including fishery resource reserves, suitable breeding areas, wind power development potential indexes, and shipping capacity, and the activation function adopts a linear activation function.

[0028] In one embodiment, the training process of the machine learning sub-model includes: The sample data corresponding to the sub-regions are divided into a training set and a validation set in a ratio of 7:3. The training batch size is set to 64, the maximum training iteration number is set to 5000 times, the Adam optimization algorithm is adopted, the initial learning rate is set to 0.001, the momentum parameter β1=0.9, the second-order momentum parameter β2=0.999, the numerical stability parameter ε=1e-8, and the mean square error (MSE) of the predicted value and the actual value is used as the loss function. During the training process, the MSE of the validation set is calculated every 100 iterations. If the MSE of the validation set does not decrease for 200 consecutive iterations, the training is stopped, and the current model parameters are saved as the trained machine learning sub-model.

[0029] Step S4, 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, the parameters of each sub-model are updated through back propagation, and the distribution of the prediction results of each sub-region machine learning sub-model is aligned.

[0030] In one embodiment, the contrastive learning loss function adopts a hybrid loss function based on the combination of KL divergence and contrastive loss, expressed as:

[0031] wherein, is a weight coefficient, exemplarily taking a value of 0.6, KL(P||Q) is the KL divergence of the prediction result distribution P and Q of the 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; D is a marginal parameter, exemplarily taking a value of 2.0; It can be understood that before calculating the KL divergence and the Euclidean distance , the prediction results can be normalized (such as Min-Max standardization, Z-Score standardization) to eliminate the dimensional difference, which will not be described here.

[0032] In one embodiment, by minimizing the , the parameters of the full connection layer and the LSTM layer of each sub-model are updated, so that the prediction distribution of the sub-models of the similar sub-regions is closer, and the prediction distribution of the sub-models of the different sub-regions maintains a reasonable degree of differentiation.

[0033] In one embodiment, exemplarily, the sub-region similarity is calculated from two dimensions of environmental characteristics (water temperature, salinity, water depth, sea current, etc.) and resource types (fish farming, wind power development, shipping, etc.). The environmental characteristic similarity is obtained by calculating the cosine similarity of the mean vector of the environmental parameters of the two sub-regions, including seawater temperature, salinity, sea current speed, water depth, etc. The resource type similarity is determined according to the main resource type (such as fish farming, wind power development, shipping, etc.) of the sub-region. If the main resource types of the two sub-regions are the same, the resource type similarity is 1, otherwise it is 0.

[0034] Specifically, exemplarily, for each sub-region, the mean vector of the environmental parameters such as seawater temperature , salinity , sea current speed , water depth , etc. is extracted, for example, the mean vector of the sub-region i is:

[0035] The cosine similarity is used to calculate the environmental characteristic similarity of the sub-regions i and j :

[0036] wherein, is the vector dot product, is the vector length, with the value range [0,1], and the value closer to 1 indicates that the environmental features are more similar.

[0037] For resource type similarity According to the core resource types of marine three-dimensional utilization (such as fishery culture, wind power development, shipping passage, and ecological protection), the main resource type of each sub-region is marked (such as sub-region i is marked as “fishery culture dominant area”). If the main resource types of sub-regions i and j are consistent, , otherwise, .

[0038] By weighted fusion of environmental features and resource type similarity, the comprehensive similarity score of sub-regions i and j is obtained:

[0039] wherein, is the weight coefficient (which can be adjusted according to the characteristics of the sea area), and exemplarily, . Exemplarily, when , it is determined that (similar sub-region), otherwise (difference sub-region).

[0040] For the KL divergence calculation of the prediction result, for example, for the corresponding sub-model (CNN-LSTM sub-model), the output prediction result vector is obtained ; Specifically, for the result vector output by the sub-model, the discrete vector value is converted into a continuous prediction result distribution by a probability density estimation method, which exemplarily includes: The result vector set output by the same sub-model under multiple groups of training / test samples , wherein m is the sample number, and the standardization processing (such as Z-Score standardization) is performed to eliminate the influence of the dimension difference of the index on the distribution estimation. The kernel density estimation algorithm (KDE) is adopted, and the pre-processed result vector set is taken as the sample to construct a continuous probability density function (i.e. the prediction result distribution ).

[0041] wherein the kernel function selects the Gaussian kernel (which satisfies the smoothness assumption of the marine evaluation index), and the bandwidth is optimized through cross-validation (such as 5-fold cross-validation) to ensure the accuracy of the distribution estimation; the obtained probability density function is taken as the prediction result distribution of the sub-model, and x is the value of the evaluation index.

[0042] Then, for each pair of similar sub-regions, the , which measures the difference between the two predictive distributions (the smaller the value, the more similar the distributions), is given by:

[0043] where, , is the probability of the sub-model predictive distribution at sample point x.

[0044] In one embodiment, for backpropagation, it updates the fully connected layer and LSTM layer parameters of the sub-model (parameters of the CNN layer are not updated, as the CNN layer is responsible for extracting general spatial features and needs to remain stable to avoid damaging the general spatial feature extraction capability; the fully connected layer and the LSTM layer are responsible for learning sub-region-specific rules and need to be aligned) as the optimization target, for example, Through gradient descent, the parameters are adjusted to minimize , which reduces the KL divergence and for similar sub-regions; and ensures for different sub-regions, maintaining the discriminability.

[0045] When the validation set no longer decreases, or the alignment of the predictive distribution of similar sub-regions (such as the average KL divergence (<0.1)) and the discriminability of different sub-regions (the proportion of > 95%) meet the requirements, the training is stopped.

[0046] Thus, through the contrastive learning loss function, the predictive distributions of similar sub-regions are aligned, and the model can transfer the rules of sample-dense sub-regions to sample-sparse sub-regions (for example, in two similar sub-regions, one is sample-dense and the other is sample-sparse). At the same time, the discriminability constraint of different sub-regions avoids confusing the characteristics of different regions. Let the model learn the common rules across regions while respecting the differences between sub-regions, ultimately improving the accuracy and robustness of the whole sea area resource evaluation and prediction.

[0047] Step S5, collect the current data of the target sea area, and input the preprocessed data into the updated machine learning sub-model of each sub-region to obtain the first prediction result of the sea area three-dimensional utilization resource evaluation.

[0048] In one embodiment, the target sea area data is collected and the collected data is preprocessed in the same way as in the previous embodiment, which will not be described here.

[0049] Subsequently, through the spatial coordinates (latitude, longitude, and water depth) of the real-time data, it is matched to the corresponding sub-region code (such as the nearshore aquaculture area, the farshore shipping area, etc., refer to the previous area division rules); the machine learning sub-model (such as the CNN-LSTM hybrid network, the model optimized by contrastive learning) dedicated to the sub-region is called. According to the region where the sub-model is located, the corresponding resource evaluation result is output.

[0050] Optionally, based on the target sea area data during historical extreme climate events, an extreme event sample library is constructed; a pre-trained large language model is used to extract features and reason data in the extreme event sample library, and a correction factor calculation model is trained in combination with actual prediction bias under extreme events. The pre-trained large language model adopts a marine field special model based on a Transformer architecture, which is obtained by pre-training on massive marine observation data and marine literature corpus. The correction factor calculation model adopts a gradient boosting decision tree algorithm, takes the feature vector output by the large language model as input, and takes the actual prediction bias as output label, and trains the correction factor calculation model.

[0051] In the present embodiment, historical extreme climate event data is collected through a multi-source data acquisition system, covering typhoon, El Nino, La Nina, strong cold wave, heavy rain, storm surge, red tide and other event types. Exemplarily, the data sources include: Marine observation network: historical monitoring data of multiple marine buoys (including anchored buoys, drifting buoys), underwater sensor arrays, and shore-based radars in the target sea area, obtaining sea water temperature, salinity, pH value, dissolved oxygen concentration, sea current speed / direction, wave height / period, tide level and other marine environmental parameters 3 months before the event, during the event and 1 month after the event, with a time resolution of 10 minutes / time and a spatial resolution of 100m x 100m.

[0052] Fishery and marine development records: historical data of coastal fishing port management system, offshore wind power operation platform, and shipping management AIS system, including fishery catch yield / operation area, wind power generation / equipment status, ship passage number / route adjustment and other human activity and resource utilization data during the event.

[0053] Weather and marine forecast archives: historical data of national meteorological center and marine forecast station, supplemented with event type (such as typhoon-strong typhoon level), intensity (typhoon wind speed, El Nino sea temperature anomaly), and influence range (latitude and longitude coverage area) annotation information.

[0054] Then, the repeated and invalid data (such as abnormal values caused by sensor failure) are removed, and the 3σ principle is used in combination with local weighted regression to correct abnormal values (such as sudden change data of sea current speed of a buoy during a typhoon, which is corrected by interpolation of surrounding normal data).

[0055] For each sample, the event type (such as "typhoon - strong typhoon level"), intensity level (typhoon is divided into tropical storm to super typhoon according to wind speed), influence range (latitude and longitude boundary), resource impact degree (fishery resource loss rate, aquaculture facility damage rate) are labeled to form a standardized labeling system. For example, a certain typhoon sample is labeled as: "event type: typhoon - strong typhoon level; intensity: wind speed 45 m / s; influence range: north latitude 30-32°, east longitude 121-123°; resource impact degree: fishery resource loss rate 25%, aquaculture facility damage rate 30%".

[0056] In this embodiment, the large language model based on the Transformer architecture, the initialization parameters are based on the general language model (such as BERT-Base), and the following marine field data is used for pre-training: Massive marine observation data: Argo buoy global marine profile data set (containing temperature, salinity profile data, more than 1 million), CMIP6 marine model simulation data set (covering different climate scenarios of marine environmental parameters, more than 500,000), providing the spatio-temporal distribution of marine environmental parameters.

[0057] Marine literature corpus: marine papers, marine industry reports collected by Web of Science, CNKI, etc., covering professional knowledge and semantic logic in the fields of marine ecology, resource development, disaster warning, etc.

[0058] Then, the model is trained, for example, including pre-training task design and training strategy. Specifically, For pre-training task design, for example: Masked language modeling (MLM): randomly mask marine environmental parameters and professional terms (such as "thermocline" "upwelling"), let the model learn the semantic association in the marine field, and predict the correct vocabulary at the mask position.

[0059] Sequence prediction task: input the time series of marine environmental parameters, let the model predict the parameter change in the next period, and learn the time series rule of marine data.

[0060] For training strategy, for example: Freeze-thaw strategy is adopted: the bottom 6 layers of the Transformer are frozen for the first 1000 iterations, only the top 6 layers are trained, the learning rate is 0.0001; all layers are unfrozen for the last 1000 iterations, the whole model is fine-tuned, the batch size is 32, and the optimizer is AdamW.

[0061] In this embodiment, the data in the extreme event sample library is input into a pre-trained large language model, and a 1024-dimensional feature vector is output, covering the environmental characteristics and semantic associations of the event (such as the potential relationship between sea temperature, sea current and fishery loss during a typhoon). Principal component analysis (PCA) is used to reduce the dimensionality of the feature vector, retaining 95% of the information, to obtain a 64-dimensional feature vector; then, the Min-Max standardization is used to map to the [0, 1] interval, eliminating the dimension effect, as the input of the correction factor calculation model.

[0062] Then, the actual prediction bias under the extreme event is calculated: the data during the event is input into the sub-regional machine learning model (such as the CNN-LSTM sub-model), to obtain the preliminary prediction value (such as the fishery resource reserve); compared with the actual resource evaluation result after the event (such as the actual catch), the difference is the actual prediction bias (label value). For example, the preliminary prediction of fishery resource reserve is 500 tons, and the actual value is 400 tons, then the bias is 100 tons.

[0063] Then, the GBDT model is trained with the feature vector output by the large language model as input and the actual prediction bias as output label to obtain the correction factor calculation model.

[0064] Specifically, for example, the number of decision trees of the GBDT model is set to 200, the maximum depth of each tree is 8, the minimum sample split number is 20, the minimum sample leaf node number is 10, the learning rate is 0.01, and the loss function is mean square error (MSE). The parameters are optimized by grid search, and the optimal parameter combination is determined by minimizing the MSE of the validation set (the sample library data is divided by 7:3).

[0065] Optionally, based on the current data of the target sea area collected in real time, 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 and output as the final sea area three-dimensional utilization resource evaluation prediction result.

[0066] In one embodiment, the wind speed, sea current speed, and sea water temperature abnormal change amplitude of the sea area are monitored in real time, and if the wind speed is greater than 12 (typhoon level), the sea current speed mutation amplitude exceeds 50%, the sea water temperature week change amplitude exceeds 3°C, or an extreme climate warning signal is received from the meteorological department, it is determined that the current is in an abnormal environmental condition; When it is determined that the target sea area is in an abnormal environmental condition, the following correction process is performed, for example: Feature vector generation: input the real-time data (including environmental parameters and event annotation information) under the abnormal environment into the Transformer model pre-trained by massive marine data, output a 1024-dimensional feature vector, and after PCA dimension reduction (retaining 95% of the information) and Min-Max standardization, a 64-dimensional input feature vector is obtained .

[0067] Correction factor prediction: the input pre-trained GBDT correction factor calculation model, the model is based on the pre-trained extreme event feature-prediction deviation mapping relationship, output correction factor .

[0068] After obtaining the first prediction result output by the sub-region machine learning sub-model, the following formula is used for correction:

[0069] wherein, is the preliminary prediction result output by the sub-model, is the final prediction result after correction. If the correction factor is negative, it means that the preliminary prediction overestimates the actual value (such as typhoon causing loss of fishery resources, which needs to adjust the prediction value); if it is positive, it means that the preliminary prediction is underestimated (such as El Nino leading to enrichment of certain fishery resources, which needs to be adjusted).

[0070] Thus, by using a pre-trained large language model to mine event features and combining a GBDT correction factor model, the preliminary prediction result is dynamically adjusted to solve the prediction deviation problem of traditional models in extreme environments and adapt to the complex needs of sea area three-dimensional utilization resource evaluation.

[0071] Figure 3 A sea area three-dimensional utilization resource evaluation prediction system 300 based on machine learning is shown. The device embodiment corresponds to the method embodiment shown in Figure 1 , specifically comprising: Target sea area data preprocessing module 301, for based on the multidimensional original data of the target sea area, sequentially performing repeated data elimination, outlier detection and correction, data standardization processing on the multidimensional original data, forming a standardized data set; Target sea area sub-region division module 302, for based on the sea area feature parameters in the standardized data set, dividing the target sea area from three dimensions of horizontal space, vertical depth and hydrodynamic, obtaining a plurality of sub-regions, the sea area feature parameters include offshore distance, water depth, current velocity, vorticity; Sub-region exclusive machine learning sub-model construction and training module 303, for each of the sub-regions, extracting sample data of the corresponding sub-region from the standardized data set, constructing a machine learning sub-model for the sub-region, training the machine learning sub-model using the sample data of the corresponding sub-region, obtaining the trained sub-region exclusive machine learning sub-model; The sub-model parameter updating module 304 based on the contrast learning loss function is used for constructing a contrast learning loss function, inputting the prediction results of all trained machine learning sub-models into the contrast learning loss function, calculating the distribution difference of the prediction results of different sub-models, updating the parameters of each sub-model through back propagation, and aligning the distribution of the prediction results of each sub-area machine learning sub-model. The sea area three-dimensional utilization resource evaluation prediction module 305 is used for collecting current data of a target sea area, inputting the preprocessed data into the updated machine learning sub-models of each sub-area, and obtaining a first prediction result of the sea area three-dimensional utilization resource evaluation.

[0072] In the above embodiments, the description of each embodiment has its own focus. If a part is not described or recorded in detail in a certain embodiment, please refer to the relevant description of other embodiments.

[0073] The above-described embodiments are only used to illustrate the technical solutions of the present application, and not to limit them. Although the present application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that they can modify the technical solutions recorded in the foregoing embodiments, or make equivalent replacements for some technical features. Such modifications or replacements do not change the essence of the corresponding technical solutions, and should be included in the protection scope of the present 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, the parameters of each sub-model are 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 first prediction result of the sea area three-dimensional utilization resource evaluation. 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 environment 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, which comprises an input layer, a CNN feature extraction layer, an LSTM time series 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 series processing layer is provided with two LSTM units, and the output layer comprises neurons consistent with the number of evaluation prediction indexes. 5.The method of claim 4, wherein, Comprise: 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.

6. The machine learning-based sea area three-dimensional utilization resource evaluation prediction method according to claim 5, 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 machine learning-based sea area three-dimensional utilization resource evaluation prediction method according to claim 6, characterized in that, The pre-trained large language model adopts a marine field special model based on a 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 a 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, and is trained to obtain the correction factor calculation model.

9. The method according to claim 4, wherein the method is characterized in that, the input features include standardized sea area geographical location features, marine environment parameter features, and human activity features; the evaluation and prediction indexes include fishery resource reserves, suitable aquaculture areas, wind power development potential indexes, and shipping capacity; and the activation function adopts a linear activation function. 10.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, 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, 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 feature parameters in the standardized data set, wherein the sea area feature parameters include offshore distance, water depth, sea current velocity, and vorticity; a sub-region exclusive machine learning sub-model construction and training module, 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, 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, configured to collect current data of the 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.

Images