A method and system for offshore wind farm site selection based on multi-source data fusion

Abstract

The present application relates to the technical field of data processing, and more particularly to a method and system for offshore wind farm site selection based on multi-source data fusion; the method can effectively solve the problem of low accuracy of offshore wind farm site selection in the prior art by constructing a multi-source spatio-temporal database, performing initial screening based on constraint conditions, and performing quantitative evaluation of industrial integration suitability, compared with the traditional single energy-oriented site selection method. A decision system including a database construction module, a candidate wind farm site selection screening module, and an offshore wind farm site selection module is constructed to promote the transformation of offshore wind power development towards high efficiency, synergy and sustainability.

Description

Technical Field

[0001] This invention relates to the field of data processing technology, and in particular to a method and system for offshore wind farm site selection decision based on multi-source data fusion. Background Technology

[0002] Offshore wind farm site selection is the primary step in offshore wind power development. Traditional methods primarily follow a technical approach centered on wind energy resources, gradually adding constraints. Specifically, traditional methods first obtain key wind energy resource parameters such as annual average wind speed, wind power density, and wind direction frequency in the target sea area through numerical simulations or measured data, screening out potential areas with acceptable wind speeds. Based on this, constraints such as water depth, geological conditions, and marine environmental protection requirements are gradually added to ultimately determine the site range for the wind farm. However, with the diversification of the marine economy, the function of offshore wind farms has shifted from single-function power generation to a comprehensive energy hub and industrial platform. Traditional site selection methods only focus on wind energy resources and basic constraints, failing to systematically consider the spatial needs and functional synergy of industrial integration during the planning stage, resulting in low efficiency in the utilization of marine space resources.

[0003] Meanwhile, existing technologies have solutions for quantitative assessment of industrial integration suitability in non-offshore wind farm site selection. However, for offshore wind farm site selection, industrial integration suitability assessment needs to process both spatial and attribute data, and there are complex nonlinear interaction relationships between the data. Commonly used improved models based on CNN models are only good at spatial feature extraction and cannot take into account the fusion of multi-source spatiotemporal data and the modeling of complex relationships in industrial integration suitability assessment, resulting in low site selection accuracy. Summary of the Invention

[0004] To address the aforementioned technical problems, this invention provides a method and system for offshore wind farm site selection based on multi-source data fusion, which solves the problems existing in the prior art.

[0005] This invention provides a method for offshore wind farm site selection decision-making based on multi-source data fusion, comprising the following steps:

[0006] S1: Construct a multi-source spatiotemporal database for offshore wind farm site selection decisions;

[0007] S2: By judging the constraints, the multi-source spatiotemporal database is initially screened to obtain multiple candidate wind farm sites;

[0008] S3: Conduct a quantitative assessment of the suitability of the multiple candidate wind farm sites for industrial integration to determine the offshore wind farm sites;

[0009] Specifically, S3 is:

[0010] S3.1: Perform data preprocessing on the multi-source spatiotemporal data corresponding to the candidate wind farm site selection;

[0011] S3.2: Construct an assessment model for quantitative evaluation of industrial integration suitability; the assessment model includes an input layer, a multimodal input and preprocessing layer, a spatial perception enhancement module, and an output layer; the spatial perception enhancement module includes orientation-sensitive convolutional blocks, a multi-scale spatial pyramid pooling module, and a spatial location encoding module;

[0012] S3.3: Train the evaluation model;

[0013] S3.4: Input the preprocessed multi-source spatiotemporal data corresponding to the candidate wind farm site selection into the evaluation model to obtain the final offshore wind farm site selection.

[0014] Preferably, the multimodal input and preprocessing layer is used to convert the initially screened multi-source spatiotemporal data into a format that the evaluation model can process, and to extract preliminary features. The multimodal input and preprocessing layer includes a spatial data processing path and an attribute data processing path. The input of the spatial data processing path is the preprocessed spatial feature data, and local spatial neighborhood features are extracted through a three-dimensional convolutional layer, with the output being a spatial feature map. The input of the attribute data path is the preprocessed attribute feature data, and the attribute feature data of each candidate wind farm site selection is mapped into a high-dimensional attribute feature vector through a fully connected layer.

[0015] Preferably, the orientation-sensitive convolutional block uses eight sets of orientation convolutional kernels covering 0°, 45°, 90°, 135°, 180°, 225°, 270°, and 315°, respectively capturing spatial feature maps in different directions and outputting eight sets of orientation feature maps, which are then combined into orientation feature vectors through channel concatenation. The size of each set of convolutional kernels is 3×3.

[0016] Preferably, the multi-scale spatial pyramid pooling module includes a dynamic scale pooling window determination module and a cross-level feature interaction module. The dynamic scale pooling window determination module dynamically adjusts the scale of the pooling window according to the input directional feature vector. The dynamic scale pooling window determination module first analyzes the input directional feature vector through a lightweight scale prediction sub-network to predict the most suitable scale index for each spatial location (i,j). The scale prediction sub-network consists of two convolutional layers and a Softmax activation layer, outputs a scale probability distribution, and selects the scale with the highest probability as the pooling window for that spatial location (i,j). The cross-level feature interaction module constructs three spatial levels in parallel for the input directional feature vector, where the first level is the original resolution of the directional feature vector, the second level is twice the original resolution, and the third level is four times the original resolution. Each level independently applies the dynamic scale pooling window for feature extraction, and then the detailed features of the first level and the semantic features of the second level are jointly transmitted to the second level through skip connections to achieve feature fusion.

[0017] Preferably, during the feature fusion process, the fusion weights of features at different levels are calculated by a hierarchical attention module, and feature fusion is achieved based on the fusion weights.

[0018] Preferably, in S2, the constraint conditions include restrictive constraint conditions and optimization constraint conditions; the restrictive constraint conditions are conditions with clear quantitative requirements for engineering feasibility, ecological safety or operational efficiency, and the optimization constraint conditions are non-rigid conditions that improve economic efficiency or operation and maintenance convenience.

[0019] Preferably, the restrictive constraints include constraints related to the natural conditions of the marine area, ecological constraints, and hard conflict avoidance constraints related to the current state of development and utilization.

[0020] Preferably, the optimized constraints include the distance to the power grid connection point, the distance to the onshore operation and maintenance base, and the synergistic potential with existing marine industries.

[0021] Preferably, the multi-source spatiotemporal data in the multi-source spatiotemporal database includes marine natural conditions data, resource endowment data, ecological constraints data, and current development and utilization data.

[0022] According to another aspect of the present invention, a multi-source data fusion-based offshore wind farm site selection decision system is provided. This system employs the aforementioned multi-source data fusion-based offshore wind farm site selection decision method. The system includes:

[0023] The database construction module builds a multi-source spatiotemporal database for offshore wind farm site selection decisions.

[0024] The candidate wind farm site selection screening module performs preliminary screening on the multi-source spatiotemporal database by judging the constraint conditions, and obtains multiple candidate wind farm sites.

[0025] The offshore wind farm site selection module performs a quantitative assessment of the suitability of multiple candidate wind farm sites for industrial integration, and determines the offshore wind farm site.

[0026] Compared with the prior art, the present invention has the following beneficial effects:

[0027] The offshore wind farm site selection decision method proposed in this invention, based on multi-source data fusion, through innovative steps such as constructing a multi-source spatiotemporal database, initial screening of constraint conditions, and quantitative assessment of industrial integration suitability, can effectively solve the problem of low accuracy in offshore wind farm site selection in existing technologies compared with traditional single-energy-oriented site selection methods, and promote the transformation of offshore wind power development towards high efficiency, collaboration and sustainability.

[0028] In constructing the evaluation model, this invention incorporates a multimodal input and preprocessing layer and a spatial perception enhancement module, based on the characteristics of wind farm site selection. The spatial perception enhancement module addresses the problem of traditional convolutional neural network models' insufficient capture of geographic spatial directionality and multi-scale features by enhancing the model's ability to perceive marine spatial features through orientation-sensitive convolution and multi-scale spatial pyramid pooling modules. The multi-scale spatial pyramid pooling module achieves accurate capture of the multi-scale dynamic characteristics of marine features through three major improvements: dynamic scale selection, cross-level feature interaction, and hierarchical weight optimization. Attached Figure Description

[0029] To more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the drawings used in the description of the specific embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0030] Figure 1 This is a flowchart of an offshore wind farm site selection decision method based on multi-source data fusion provided by an embodiment of the present invention;

[0031] Figure 2 This is a flowchart provided by an embodiment of the present invention for quantitatively assessing the suitability of multiple candidate wind farm sites for industrial integration and determining the site selection for offshore wind farms. Detailed Implementation

[0032] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below. Obviously, the described embodiments are only a part of the embodiments of this invention, and not all of them. Based on the embodiments of this invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this invention.

[0033] Example 1

[0034] like Figure 1 As shown, this invention proposes a method for offshore wind farm site selection decision-making based on multi-source data fusion, including the following steps:

[0035] S1: Construct a multi-source spatiotemporal database for offshore wind farm site selection decisions;

[0036] In this step, the multi-source spatiotemporal database integrates all key data affecting offshore wind power site selection and forms a comprehensive, high-precision, and dynamically updated database covering natural conditions, resource endowment, ecological constraints, and current development status through spatial geographic coordinate association.

[0037] The multi-source spatiotemporal database includes data on marine natural conditions, resource endowment, ecological constraints, and current development and utilization status. The marine natural conditions data describes the physical environmental characteristics of the sea area, directly impacting the structural design, foundation type, construction difficulty, and long-term operation and maintenance costs of wind turbines. This data includes water depth data, seabed topography and geological data, ocean current and tidal data, and extreme weather and marine environment data. The water depth data represents the vertical depth from sea level to the seabed. The seabed topography data is derived from digital elevation models to describe seabed undulations. The geological data includes seismic intensity, soil and rock types, bearing capacity, and fault distribution. The ocean current and tidal data includes current velocity and direction. The extreme weather and marine environment data includes the path of a 50-year typhoon, maximum wave height, and maximum wind speed. The data includes wave height, wave period, etc.; the resource endowment data includes annual average wind speed, wind power density, and wind direction frequency; the ecological constraint data includes marine protected area data, sensitive ecological habitat data, and other ecologically related data. The marine protected area data includes the boundary coordinates and functional zoning data of the protected area. The sensitive ecological habitat data includes the distribution range and coverage of coral reefs / seagrass beds, the water depth of coral reefs, the distribution boundary of mangrove wetlands, tree species types, and coordinates of the activity range of rare species, etc.; other ecologically related data includes the coordinates of the sea areas where juvenile fish concentrate to breed, etc.; the current development and utilization status data is used to describe the current and planned development and utilization of the sea area to avoid functional conflicts between wind farms and existing facilities. The current development and utilization status data includes the main channels for ships to enter and leave the port, the center line of the marked channel, the safety buffer zone, and the temporary berthing area for ships, etc.

[0038] In this step, when constructing the multi-source spatiotemporal database, data of different formats are converted to a format supported by the GeoPackage database. All data are converted to the Chinese CGCS2000 coordinate system and UTM projection method to ensure that spatial locations can be overlaid for analysis.

[0039] The multi-source spatiotemporal database is constructed using PostgreSQL and the spatial extension module PostGIS, and the data is divided into the following core tables:

[0040] Basic geographic table: including marine boundaries and coordinate reference system definitions;

[0041] Natural conditions table: water depth, topography, geological parameters, ocean currents;

[0042] Resource endowment table: Wind energy resource data;

[0043] Ecological constraints: protected areas, sensitive habitats, and areas where rare species may live;

[0044] Current Development Status Table: Waterway.

[0045] The multi-source spatiotemporal database associates data from different tables with unique identifiers. For example, it can bind wind energy resource data at a certain coordinate point with water depth and ecological constraint data at that location to form a structure with multiple attributes at a single point.

[0046] S2: By judging the constraints, the multi-source spatiotemporal database is initially screened to obtain multiple candidate wind farm sites;

[0047] The constraints include restrictive constraints and optimization constraints; the restrictive constraints are those with clear quantitative requirements for project feasibility, ecological safety, or operational efficiency, while the optimization constraints are non-rigid conditions that improve economic efficiency or ease of operation and maintenance.

[0048] The restrictive constraints require multi-condition joint screening to ensure that the candidate wind farm sites simultaneously meet all restrictive requirements. These constraints include constraints related to marine natural conditions, ecological conditions, and the avoidance of hard conflicts arising from the current state of development and utilization. Marine natural conditions constraints include: preference for shallow sea areas over deep sea areas; sufficient support from seabed rock and soil for wind turbine foundations; an average annual ocean current velocity ≤2 m / s; and avoiding areas where extreme wind speeds / wave heights during a 50-year typhoon exceed design thresholds. Ecological constraints include meeting minimum distance requirements from marine protected areas and areas frequented by rare species. The avoidance of hard conflicts arising from the current state of development and utilization includes ensuring a minimum distance between the wind farm and the centerline of shipping lanes for safe navigation.

[0049] Although the optimized constraints do not directly determine whether a site is qualified, they can affect subsequent development costs or efficiency and can be used as a preferred bonus in the initial screening. The optimized constraints include the distance to the power grid connection point, the distance to the onshore operation and maintenance base, and the synergistic potential with the existing marine industry.

[0050] The restrictive and optimized constraints can be used to initially screen the target area and obtain multiple candidate wind farm sites.

[0051] S3: Conduct a quantitative assessment of the suitability of the multiple candidate wind farm sites for industrial integration to determine the offshore wind farm sites;

[0052] Current offshore wind farm site selection primarily focuses on wind energy resources themselves, neglecting the potential for multi-purpose development of the vast sea areas occupied by wind farms. For example, the spacing between wind turbine foundations is typically 5-10 times the rotor diameter, leaving ample space for other industrial development; ancillary facilities such as submarine cable corridors and maintenance access routes can also be shared with adjacent industries. Existing technologies face challenges in quantitatively assessing the suitability for industrial integration, including difficulties in fusing multi-source heterogeneous data, insufficient modeling of nonlinear relationships, and a lack of interpretability.

[0053] This embodiment proposes a deep learning model for multimodal spatial attention, and improves its structure to meet the characteristics of industrial integration suitability assessment. By adding a specific spatial perception module, it achieves accurate assessment of complex industrial integration scenarios.

[0054] like Figure 2 As shown, S3 specifically includes:

[0055] S3.1: Perform data preprocessing on the multi-source spatiotemporal data corresponding to the candidate wind farm site selection;

[0056] The data preprocessing includes data classification, spatial data rasterization and alignment, and feature encoding.

[0057] Specifically, the data classification involves dividing the multi-source spatiotemporal data corresponding to the candidate wind farm sites into two main categories:

[0058] Spatial characteristic data: including wind farm location coordinates, water depth data, seabed topography data, spatial distance from ecological protection areas, spatial distance from shipping channels, etc.

[0059] Attribute characteristic data: including numerical variables such as annual average wind speed, wind power density, average water depth, geological bearing capacity, ocean current velocity, distance to power grid connection point, and distance to onshore operation and maintenance base.

[0060] The spatial data rasterization and alignment operation specifically involves resampling the spatial feature data to a regular raster grid with a resolution of 100m×100m; using a geographic coordinate system (CGCS2000) to ensure accurate spatial location matching; and normalizing each raster layer to eliminate the influence of dimensions.

[0061] The feature encoding is specifically as follows: one-hot encoding is used for categorical variables; and Min-Max standardization is used to the [0,1] interval for numerical variables.

[0062] S3.2: Construct an assessment model for quantitative evaluation of the suitability of industrial integration;

[0063] Industry integration suitability assessment requires the simultaneous processing of spatial and attribute data, with complex nonlinear interactions between these data. Traditional deep learning models, such as CNNs which excel at spatial feature extraction and MLPs which only handle tabular data, struggle to address the fusion of multi-source spatiotemporal data and the modeling of complex relationships in industry integration suitability assessment. In this step, the assessment model includes an input layer, a multimodal input and preprocessing layer, a spatial awareness enhancement module, and an output layer.

[0064] The input layer is used to input the preprocessed multi-source spatiotemporal data corresponding to the candidate wind farm site selection.

[0065] The multimodal input and preprocessing layer is used to convert the initially screened multi-source spatiotemporal data into a format that the evaluation model can process, and extract preliminary features. The multimodal input and preprocessing layer includes a spatial data processing path and an attribute data processing path. The input of the spatial data processing path is the preprocessed spatial feature data, and local spatial neighborhood features are extracted through a three-dimensional convolutional layer to output a spatial feature map. The input of the attribute data path is the preprocessed attribute feature data, and the attribute feature data of each candidate wind farm site selection is mapped into a high-dimensional attribute feature vector through a fully connected layer.

[0066] The spatial perception enhancement module addresses the shortcomings of traditional convolutional neural network models in capturing geographic spatial directionality and multi-scale features. It designs direction-sensitive convolution and multi-scale pyramid pooling to enhance the model's ability to perceive marine spatial features. The spatial perception enhancement module includes a direction-sensitive convolution block, a multi-scale spatial pyramid pooling module, and a spatial location encoding module. The direction-sensitive convolution block uses eight sets of directional convolution kernels covering 0°, 45°, 90°, 135°, 180°, 225°, 270°, and 315°, capturing spatial feature maps in different directions and outputting eight sets of directional feature maps. These maps are then concatenated into a directional feature vector through channel concatenation. Each set of convolution kernels is 3×3 in size to strengthen the evaluation model's recognition of spatial directional patterns.

[0067] Furthermore, the eight directional convolution kernels of the directional sensitive convolution block all adopt a separable convolution structure. Each directional convolution kernel only responds to spatial features within a specific angle range. For example, the 0° directional convolution kernel focuses on capturing the wind field distribution features in the east-west direction (along the coastline), while the 45° directional convolution kernel captures the ocean current direction features in the northeast-southwest direction.

[0068] The spatial feature maps in different directions are captured separately and eight sets of directional feature maps are output. They are then combined into a directional feature vector by channel concatenation. Specifically, the eight sets of directional feature maps are concatenated according to the channel dimension to form a fused feature map with a total of 256 channels. Then, the number of channels of the fused feature map is compressed from 256 to 64 by using a 1×1 convolution kernel to reduce the amount of computation while retaining key directional features.

[0069] Key characteristics such as the impact range of wind turbine layout and the ecological impact radius of marine ranches in the surrounding waters of offshore wind farms have dynamically changing spatial scales. Fixed-scale pooling windows are unable to adaptively capture these dynamic characteristics. Traditional multi-level spatial pyramid pooling only performs multi-scale processing at a single spatial level, ignoring the hierarchical relationships between local wind turbine disturbances and regional ocean current patterns and macro-ocean topology, resulting in limited modeling capabilities for complex spatial dependencies. Therefore, the multi-scale spatial pyramid pooling module in this embodiment achieves accurate capture of the multi-scale dynamic characteristics of marine features through three major improvements: dynamic scale selection, cross-level feature interaction, and hierarchical weight optimization.

[0070] The multi-scale spatial pyramid pooling module includes a dynamic scale pooling window determination module and a cross-level feature interaction module. The dynamic scale pooling window determination module dynamically adjusts the scale of the pooling window based on the input directional feature vector. First, it analyzes the input directional feature vector through a lightweight scale prediction sub-network to predict the most suitable scale index for each spatial location (i,j). This scale prediction sub-network consists of two convolutional layers and a Softmax activation layer, outputting a scale probability distribution and selecting the scale with the highest probability as the pooling window for that spatial location (i,j). The cross-level feature interaction module uses inter-level features... The feature fusion process establishes a multi-level spatial association between local, regional, and global features. The cross-level feature interaction module constructs three spatial levels in parallel on the input directional feature vector. The first level is the original resolution of the directional feature vector, the second level is twice the original resolution, and the third level is four times the original resolution. Each level independently applies a dynamic scale pooling window for feature extraction. Then, through a skip connection, the detailed features of the first level and the semantic features of the second level are jointly transmitted to the second level to achieve feature fusion. During the feature fusion process, the level attention module (LAM) calculates the fusion weights of features at different levels, and feature fusion is achieved based on the fusion weights.

[0071] The hierarchical attention module includes a 1×1 convolutional layer, a global average pooling layer, a dot product attention mechanism layer, and a Softmax function layer. The process of calculating the fusion weights for different hierarchical features in the hierarchical attention module is as follows: Each hierarchical feature is compressed in number of channels using the 1×1 convolutional layer, reducing each feature from 64 channels to 32 channels, and a query vector is generated; then, the global average pooling layer performs global average pooling on all hierarchical features to generate a global feature vector; the dot product attention mechanism layer calculates the similarity between the query vector and the global context vector using dot product operations to obtain an initial score for the hierarchical feature and the global context vector; finally, the Softmax function layer converts the initial score into fusion weights.

[0072] In this step, dynamic scale selection and cross-level interaction enable the evaluation model to adapt to the multi-scale characteristics of different sea areas, thereby improving the evaluation model's ability to model complex spatial dependencies.

[0073] The output layer is used to map the fused features into an industry integration suitability score, and it consists of a fully connected layer and a Sigmoid activation function.

[0074] S3.3: Train the evaluation model;

[0075] In this step, the loss function for evaluating model training includes a primary loss function and an auxiliary loss function. The primary loss function uses Huber loss instead of the traditional mean squared error (MSE) to balance outlier sensitivity. The auxiliary loss function incorporates feature importance consistency loss (L_FIC) to constrain the model to maintain the relative stability of feature importance ranking in different training batches, thereby improving model interpretability.

[0076] During training, a cosine annealing with warm restarts strategy is adopted. In the early stage, a larger learning rate is used to achieve rapid convergence, and in the later stage, the learning rate is gradually reduced for fine-tuning.

[0077] S3.4: Input the preprocessed multi-source spatiotemporal data corresponding to the candidate wind farm site selection into the evaluation model to obtain the final offshore wind farm site selection.

[0078] Example 2

[0079] This invention also provides an offshore wind farm site selection decision system based on multi-source data fusion, which adopts the offshore wind farm site selection decision method based on multi-source data fusion in Embodiment 1. The system includes:

[0080] The database construction module builds a multi-source spatiotemporal database for offshore wind farm site selection decisions.

[0081] The candidate wind farm site selection screening module performs preliminary screening on the multi-source spatiotemporal database by judging the constraint conditions, and obtains multiple candidate wind farm sites.

[0082] The offshore wind farm site selection module performs a quantitative assessment of the suitability of multiple candidate wind farm sites for industrial integration, and determines the offshore wind farm site.

[0083] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention 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 or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the technical solutions of the embodiments of the present invention.

Claims

1. A method for offshore wind farm site selection decision based on multi-source data fusion, characterized in that: Includes the following steps: S1: Construct a multi-source spatiotemporal database for offshore wind farm site selection decisions; S2: By judging the constraints, the multi-source spatiotemporal database is initially screened to obtain multiple candidate wind farm sites; S3: Conduct a quantitative assessment of the suitability of the multiple candidate wind farm sites for industrial integration to determine the offshore wind farm sites; S3 specifically comprises: S3.1: Preprocessing the multi-source spatiotemporal data corresponding to the candidate wind farm sites; S3.2: Constructing an evaluation model for quantitative assessment of industrial integration suitability; the evaluation model includes an input layer, a multimodal input and preprocessing layer, a spatial perception enhancement module, and an output layer; the spatial perception enhancement module includes a direction-sensitive convolutional block, a multi-scale spatial pyramid pooling module, and a spatial location encoding module; S3.3: Training the evaluation model; S3.4: Inputting the preprocessed multi-source spatiotemporal data corresponding to the candidate wind farm sites into the evaluation model to obtain the final offshore wind farm site selection; The data preprocessing also includes data classification, specifically dividing the multi-source spatiotemporal data corresponding to the candidate wind farm sites into spatial feature data and attribute feature data. The spatial feature data includes wind farm location coordinates, water depth data, seabed topography data, spatial distance to ecological protection zones, and spatial distance to waterways. The attribute feature data includes annual average wind speed, wind power density, average water depth, geological bearing capacity, ocean current velocity, distance to grid connection point, and distance to onshore operation and maintenance base. The multimodal input and preprocessing layer is used to convert the initially screened multi-source spatiotemporal data into a format that the evaluation model can process, and to extract preliminary features. The multimodal input and preprocessing layer includes a spatial data processing path and an attribute data processing path. The input of the spatial data processing path is the preprocessed spatial feature data, and local spatial neighborhood features are extracted through a three-dimensional convolutional layer, with the output being a spatial feature map. The input of the attribute data path is the preprocessed attribute feature data, and the attribute feature data of each candidate wind farm site selection is mapped into a high-dimensional attribute feature vector through a fully connected layer.

2. The offshore wind farm site selection decision method based on multi-source data fusion according to claim 1, characterized in that: The orientation-sensitive convolutional block uses eight sets of orientation convolutional kernels covering 0°, 45°, 90°, 135°, 180°, 225°, 270°, and 315°. These kernels capture spatial feature maps in different directions and output eight sets of orientation feature maps. The kernels are then combined into an orientation feature vector through channel concatenation. Each set of convolutional kernels is 3×3 in size.

3. The offshore wind farm site selection decision method based on multi-source data fusion according to claim 2, characterized in that: The multi-scale spatial pyramid pooling module includes a dynamic scale pooling window determination module and a cross-level feature interaction module. The dynamic scale pooling window determination module dynamically adjusts the scale of the pooling window based on the input directional feature vector. First, it analyzes the input directional feature vector through a lightweight scale prediction sub-network to predict the most suitable scale index for each spatial location (i,j). The scale prediction sub-network consists of two convolutional layers and a Softmax activation layer, outputting a scale probability distribution and selecting the scale with the highest probability as the pooling window for that spatial location (i,j). The cross-level feature interaction module constructs three spatial levels in parallel for the input directional feature vector. The first level is the original resolution of the directional feature vector, the second level is twice the original resolution, and the third level is four times the original resolution. Each level independently applies the dynamic scale pooling window for feature extraction. Then, through skip connections, the detailed features of the first level and the semantic features of the second level are jointly transmitted to the second level, achieving feature fusion.

4. The offshore wind farm site selection decision method based on multi-source data fusion according to claim 3, characterized in that: During feature fusion, the fusion weights of features at different levels are calculated through a hierarchical attention module, and feature fusion is achieved based on the fusion weights.

5. The offshore wind farm site selection decision method based on multi-source data fusion according to claim 1, characterized in that: In S2, the constraints include restrictive constraints and optimization constraints; the restrictive constraints are conditions with clear quantitative requirements for project feasibility, ecological safety or operational efficiency, and the optimization constraints are non-rigid conditions that improve economic efficiency or ease of operation and maintenance.

6. The offshore wind farm site selection decision method based on multi-source data fusion according to claim 5, characterized in that: The restrictive constraints include constraints related to the natural conditions of the marine area, ecological constraints, and hard conflict avoidance constraints related to the current state of development and utilization.

7. The offshore wind farm site selection decision method based on multi-source data fusion according to claim 5, characterized in that: The optimized constraints include the distance to the power grid connection point, the distance to the onshore operation and maintenance base, and the synergistic potential with existing marine industries.

8. The offshore wind farm site selection decision method based on multi-source data fusion according to claim 1, characterized in that: The multi-source spatiotemporal database contains multi-source spatiotemporal data including marine natural conditions data, resource endowment data, ecological constraints data, and current development and utilization data.

9. A decision-making system for offshore wind farm site selection based on multi-source data fusion, characterized in that, The system employs a multi-source data fusion-based offshore wind farm site selection decision method as described in any one of claims 1-8, and the system comprises: The database construction module builds a multi-source spatiotemporal database for offshore wind farm site selection decisions. The candidate wind farm site selection screening module performs preliminary screening on the multi-source spatiotemporal database by judging the constraint conditions, and obtains multiple candidate wind farm sites. The offshore wind farm site selection module performs a quantitative assessment of the suitability of multiple candidate wind farm sites for industrial integration, and determines the offshore wind farm site.

Images