Spatial perception type modular wind resource perception method suitable for offshore wind plant

Abstract

The invention discloses a space sensing type modular wind resource sensing method suitable for an offshore wind plant. The method comprises the following steps: step 1, arranging a wind plant three-dimensional sensing system; 2, carrying out multi-source data acquisition and edge preprocessing; 3, carrying out spatial-temporal feature modeling and intelligent prediction; and 4, wind energy evaluation and system feedback optimization are carried out. The invention provides a multi-source layout strategy combining a tower drum, a jacket, a booster station and a buoy, and a three-dimensional sensing network covering the vertical wind speed gradient, boundary layer disturbance and the overall wind field dominant trend is formed by fully utilizing the structure positions of different elevations and regions in a wind power plant. Compared with a traditional single-point or linear layout mode, the scheme realizes higher spatial resolution and wind field state restoration capability on the premise of not increasing a large amount of hardware investment, and provides a solid data foundation for high-precision prediction.

Description

Technical Field

[0001] This invention relates to the field of offshore wind power, specifically to a spatially perceptive modular wind resource sensing method applicable to offshore wind farms. Background Technology

[0002] Currently, offshore wind power, as an important form of renewable energy, is expanding in scale, placing higher demands on wind resource acquisition and forecasting technologies. Accurate acquisition of meteorological parameters such as wind speed and direction is fundamental for wind farm site selection, turbine layout, power forecasting, and operation scheduling. However, existing wind resource monitoring and forecasting systems generally suffer from limitations in structural deployment, insufficient spatial sensing capabilities, and significant data processing delays.

[0003] Common wind speed monitoring methods include installing single-point anemometers, which are simple in structure and low in cost, but cannot reflect the overall spatial distribution characteristics of the wind field. While lidar systems have some vertical scanning capability, they are expensive, consume a lot of power, and are difficult to maintain, making them unsuitable for large-scale deployment. Furthermore, although marine meteorological buoys can perform localized data collection, their low deployment density, weak anti-drift capability, and susceptibility to sea state interference result in poor data continuity.

[0004] Regarding forecasting methods, some wind farms rely on historical SCADA data to build time-series models for wind speed prediction. However, such methods lack spatial information support and struggle to cope with sudden wind changes and deployments in new areas. Systems generally rely on a centralized platform for unified calculations, lacking edge awareness and distributed processing capabilities, resulting in insufficient real-time performance and stability. Furthermore, most equipment has low structural integration, hindering modular maintenance and efficient replacement, making it difficult to meet the long-term operational needs of complex offshore environments such as high humidity, high salinity, and high vibration.

[0005] Therefore, there is an urgent need for a new type of wind resource system with a modular structure, supporting rapid multi-point deployment, edge computing capabilities, and the ability to realize wind field spatial perception and wind speed trend prediction, in order to solve the shortcomings of existing technologies in data acquisition and deployment, spatial modeling, and system maintenance. Summary of the Invention

[0006] The purpose of this invention is to provide a spatially perceptive modular wind resource sensing method suitable for offshore wind farms, in order to solve the problems mentioned in the background art.

[0007] To achieve the above objectives, the present invention provides the following technical solution: A spatially-aware, modular wind resource sensing method suitable for offshore wind farms includes: Step 1, Deploy a three-dimensional wind field sensing system, including: Step 1.1, Node selection; Step 1.2: Perform spatial site selection and three-dimensional layout optimization for wind turbine deployment nodes; Step 2 involves multi-source data acquisition and edge preprocessing, including: Step 2.1: Configure multiple sensors; Step 2.2, Edge device platform construction and task allocation; Step 2.3: Construct a data synchronization mechanism and time alignment strategy; Step 3, perform spatiotemporal feature modeling and intelligent prediction, including: Step 3.1: Construct the dataset and organize tensors; Step 3.2: Design the prediction model structure; Step 3.3: Perform physical enhancement modeling and construct a prediction correction mechanism; Step 3.4: Construct an online update and model adaptation mechanism; Step 4, conduct wind energy assessment and system feedback optimization, including: Step 4.1: Conduct a wind energy resource distribution assessment; Step 4.2: Construct a closed-loop feedback mechanism for prediction accuracy; Step 4.3: Perform wind power prediction and scheduling optimization; Step 4.4: Construct a multi-point dynamic adjustment mechanism.

[0008] Further, step 1.1 includes: Deploy data collection modules for the following five types of locations: a. Middle and top sections of the wind turbine tower; b. Edge of the jacket foundation platform; c. Top of the booster station / roof of the control center; d. Submarine cable outgoing platform and edge channel platform; e. Buoy array deployment.

[0009] Further, step 1.2 includes: The spatial location and three-dimensional layout of wind turbine deployment nodes are optimized based on the following four optimization principles: (1) Structural topology adaptability principle: Based on the geometric distribution and installability of wind turbine towers, jacket platforms, substation roofs, and buoy structures within the wind farm, a structural adaptability function is constructed. It is constructed from the structural topology diagram, with deployable areas assigned a value of 1 and non-deployable areas assigned a value of 0. Intermediate values ​​are approximated by the platform surface shape function. (2) Dominant wind direction coupling principle: Analyze the dominant wind direction of the offshore wind field. Calculate the angle between the line connecting the candidate points and the target point to obtain the prevailing wind direction response function. : In the formula, r represents the spatial position vector of any point within the offshore wind farm, r0 is the spatial vector of the reference position, and r0 is the wind vector corresponding to the prevailing wind direction. Let it be its unit vector. Indicates the angle between the prevailing wind direction and the line connecting them; (3) Wind power density driving principle: Construct wind power density P based on historical wind speeds (r) To identify areas with high wind energy resources and areas of drastic fluctuation: In the formula, The air density is represented by 'r'; the average wind speed at location 'r' is represented by 'v'. (4) Comprehensive suitability scoring principle: All locations are jointly optimized using simulated wind field heat maps, terrain shading maps, and historical wind power density data to calculate the deployment density function D(r): In the formula, max represents taking the maximum value, and w s (r) represents the weighting function for the influence of terrain and shading, w w (r) represents the dominant wind direction coupled response function; α1, α2, and α3 represent weighting coefficients; All candidate deployment nodes are assigned priority scores using the deployment density function D(r), and a deployment optimization objective function is constructed to maximize global deployment efficiency under constraints. N optimal deployment nodes are selected from the set of all candidate deployment nodes, denoted as . r1 represents the first candidate deployment node, r2 represents the second candidate deployment node, M is the total number of candidate deployment nodes, and the deployment node selection variable is... , indicating whether to select the first For each point, the following objective function is used: In the formula, D(r) i ) represents the overall suitability function, N represents the optimal number of nodes to be selected, and d min This indicates the minimum safe distance between two points; Using an integer linear programming algorithm, N optimal nodes are selected under the constraint that "the distance between any two nodes is not less than the preset minimum distance threshold", thus forming the actual engineering layout diagram.

[0010] Furthermore, in step 2.1, the following multiple sensors are configured: a three-dimensional ultrasonic anemometer, a humidity, temperature and pressure module, a radiation / visibility module, an IMU attitude module, and a power supply and status monitoring module; Step 2.2 includes: Each node is equipped with a lightweight edge computing unit, embedding the following task modules: (1) Signal acquisition and buffering Data from various sensors is acquired via serial port / I2C / analog channel and stored in a local memory queue according to timestamps, with the buffer size dynamically adjusted. (2) Anomaly removal and redundancy reduction Median filtering and first-order difference methods are used to remove abrupt transitions, and sliding window convolution compression is applied to the time series. In the formula, x filt (t) represents the smoothing result at time t, where t is the time index. This represents the data of the i-th sampling point, where w is the adjustable window width threshold; (3) Local feature extraction and preliminary clustering Wavelet analysis was performed on the wind speed sequence to extract multi-scale disturbance energy indices. In the formula, E j (t) represents the local energy at the j-th scale, and the wavelet coefficients at scale j and time i after wavelet decomposition. This represents the instantaneous energy density at the j-th scale layer number; At the same time, K-means is used to classify the degree of perturbation of time-period features; (4) Uplink preprocessing and cache backup The processed data is divided into "main feature value + anomaly record + original summary" and uploaded to the host computer or monitoring center via wireless communication module at a set frequency.

[0011] Further, step 2.3 includes: The wind turbine deployment nodes have built-in clock modules. The wind resource sensing system synchronizes its time according to a set cycle via broadcast time from the main control time of the booster station. The maximum error is controlled within a set threshold range based on project requirements. The following strategy is used for aligning the timestamps of the collected data: Let the time deviation of each node be . The uploaded data will be mapped uniformly as follows: In the formula, For the first The calibrated time of each node This is the original timestamp. This refers to the node clock offset; Multi-point data, after time alignment, are fused using a unified time base to construct a 4D tensor of the wind field: In the formula, This represents the wind field in spatial coordinates. The overall state tensor over time t, where u, v, and w are the components of wind speed in the three directions, and T, H, and P are the temperature field, humidity field, and air pressure field, respectively; the subscript t indicates the discrete time frame index.

[0012] Further, step 3.1 includes: All data reported by the optimally deployed nodes are time-aligned and spatially reconstructed by the central platform to construct the spatiotemporal tensor of the entire wind field. The specific steps are as follows: Step 3.1.1, Timing Window Construction Given a time step Δt, with a sliding window length T win Constructing a time-slice sequence X with step size ΔT k : In the formula, X k Let t represent the set of wind field tensor sequences within the k-th time window. k This indicates the end time of the k-th time window, where the subscript i is the time step index within the window, and k is the current sliding window number; Step 3.1.2, Spatial Interpolation and Mesh Unification The data of each optimally positioned node are interpolated to a regular 3D grid, using either weighted nearest neighbor interpolation or a Gaussian kernel method. in, Represents the reconstructed wind field tensor under a regular grid; This represents the observation data of the i-th node; N is the number of valid observation nodes; For interpolation weights, satisfying ; Step 3.1.3, Data Normalization and Detrending Normalization using mean-standard deviation: in, This represents the normalized signal; These are the original time series values; This represents the sample mean; This represents the sample standard deviation and is used to fit a trend term to historical daily periodic data. .

[0013] Further, step 3.2 includes: The central platform integrates the following two types of prediction models, combining the spatiotemporal structure of the wind field with historical evolution patterns, to perform multi-step predictions of wind speed and direction: (1) LSTM network based on double-layer recurrent structure The input to this LSTM network based on a two-layer recurrent structure is a normalized multidimensional time series tensor. ; The structure of this LSTM network based on a two-layer recurrent structure includes: two stacked LSTM layers; The hidden dimension of this LSTM network based on a two-layer recurrent structure is H=128, and it outputs a wind speed prediction sequence: , in, This represents the predicted future T-step wind speed sequence; This represents the normalized multidimensional input tensor, containing features such as wind speed, temperature, and humidity; LSTM1 and LSTM2 represent the first and second long short-term memory units, respectively; t is the current time index; and T is the prediction step size. (2) Lightweight Transformer Network The input encoder of this lightweight Transformer network uses a concatenation of temporal position encoding PE(t) and multidimensional feature encoding. In the formula, E in The input matrix of the Transformer network is composed of temporal location encoding and feature vectors. Represents the normalized feature tensor of the current time step; PE(t) is the temporal position code used to preserve temporal position information; The core structure of this lightweight Transformer network consists of three attention modules that use causal masks to control the direction of information flow. The output module of this lightweight Transformer network: a linear mapping outputs the future T. pred Wind speed and direction at each time step: , In the formula, Transformer(·) represents the main function of the self-attention network, which includes 3 attention modules and a feedforward layer; (3) Model selection strategy: A lightweight Transformer network is used for regions with drastic fluctuations; an LSTM network based on a two-layer recurrent structure is used for regions with stable conditions.

[0014] Further, step 3.3 includes: Introduce the physical driving factor of wind power and the topographic influence function as priors: (1) Wind energy density driver: In the formula, P wind (t) represents the wind energy power density at time t; Where A is the air density; A is the swept area of ​​the wind turbine. Average wind speed; As auxiliary feature inputs, LSTM networks or lightweight Transformer networks with a two-layer recurrent structure are used to guide the model network to focus on the nonlinear characteristics of the influence of wind speed on energy. (2) Terrain shading function: in, S is the terrain shading correction factor. obst The projected area of ​​the obstacle is used to correct the wind speed prediction bias; S total This represents the total sampling area of ​​the region; (3) Fusion correction formula: in, This represents the corrected final wind speed forecast. This indicates that the model predicts wind speed. This is a dynamic compensation term based on local historical errors; Step 3.4 includes: The system establishes an online evaluation and model self-learning mechanism, comprising the following three core modules: (1) Error monitoring module: Calculates the rolling MAPE index: In the formula, MAPE k This represents the moving average absolute percentage error for the k-th monitoring period; T is the number of samples in the current monitoring period. This indicates the model's predicted wind speed; This indicates the measured wind speed; (2) Triggering mechanism: When the model error exceeds the preset error threshold for multiple consecutive periods. Automatically triggers model adaptive update operation: (3) Adaptive update operation: Online incremental fine-tuning: Locally updating the weights of the wind resource prediction model based on the latest data samples; Remotely distribute wind resource prediction model parameters: The main control center sends the corrected model weights to each optimal deployment node; Abnormal node restart and recalibration commands: For deployed nodes with significantly excessive errors, resampling and time synchronization commands are automatically issued to restore synchronization performance.

[0015] Further, step 4.1 includes: Based on the predicted wind speed sequences of each optimal deployment node Calculate the local wind energy density distribution to generate a wind farm heat map, which can be used for spatial assessment and power dispatching reference. in, air density; Represents spatial coordinates Wind speed sequence over time t; Wind energy density per unit area; The total instantaneous wind energy P can be estimated by performing three-dimensional integration over the entire deployment area. total (t): The wind resource sensing system generates the following based on this: regional wind energy heat map, time-series wind energy trend map, and wind energy output prediction curve. Step 4.2 includes: The wind resource sensing system constructs a closed-loop feedback channel based on prediction errors and actual measurement errors to achieve adaptive fine-tuning of the wind resource prediction network model. The core indicator is the weighted relative error. Among them, WRE i This represents the weighted relative error index; w i Point energy weighting factor: When the average WRE within a continuous window exceeds the preset error threshold At that time, the wind resource forecasting system triggers the following feedback strategy: The weights of the wind speed prediction model are incrementally updated based on the latest observation data to achieve self-learning and accuracy correction during operation; when the error of a specific measuring point continues to exceed the limit, parameter adjustment suggestions are automatically generated; when the wind energy in a local area changes drastically or the boundary energy gradient is significant, the wind resource prediction system generates deployment optimization suggestions.

[0016] Further, step 4.3 includes: The wind speed forecast is converted into a wind turbine output estimate, taking into account the wind turbine power curve P(u), to calculate the wind power output. predict: in, This represents the predicted total output power at time t. This represents the power value corresponding to the power curve of the i-th wind turbine; represents the operating efficiency parameter of the i-th wind turbine; N is the total number of wind turbines; the prediction results are used to guide short-term wind energy dispatch, charging and discharging energy storage strategies, and grid connection curve correction. Step 4.4 includes: The wind resource forecasting system supports a periodic optimization and addition / removal mechanism for deployed nodes, which utilizes a dynamic evaluation function. A comprehensive evaluation of the operational status and contribution of each deployed node is conducted to achieve adaptive optimization and adjustment of the wind field sensing network. in, This represents the overall performance score of the i-th deployed node. This represents the information gain of the i-th deployment node; This represents the signal stability index of the i-th deployment node; Indicates the first The layout structure value of each deployment node; α, β, and γ represent weight coefficients, which are set according to the system optimization objective; Wind resource forecasting system based on Dynamic trend analysis and automatic identification of efficient and inefficient nodes: When the node performance score is lower than the preset threshold θ, the wind resource prediction system automatically marks "inefficient points" and provides suggestions for layout adjustment, node status adjustment, and new node optimization to maintain the optimal state of the overall wind field perception capability of the wind resource prediction system; newly added points are reselected according to the node deployment optimization function in step 1 to keep the overall wind field perception capability optimal.

[0017] Compared with the prior art, the beneficial effects of the present invention are:

[0018] 1) This invention proposes a multi-source deployment strategy combining wind turbine towers, jacket structures, substations, and buoys. This strategy fully utilizes the structural locations at different elevations and in different areas within the wind farm to form a three-dimensional sensing network encompassing vertical wind speed gradients, boundary layer disturbances, and the overall dominant wind field trend. Compared to traditional single-point or linear deployment methods, this approach achieves higher spatial resolution and wind field state reconstruction capabilities without significantly increasing hardware investment, providing a solid data foundation for high-precision prediction.

[0019] 2) Each acquisition node has a built-in edge computing module, which has basic filtering, data compression, and outlier removal functions. The system extracts key feature data locally and uploads it in a structured format, effectively reducing communication bandwidth consumption and avoiding network congestion caused by large-scale backhaul of raw data; at the same time, it provides a cleaner and more efficient input data stream for the central platform's time series modeling and multi-point fusion prediction. Attached Figure Description

[0020] Figure 1 This is a flowchart of a spatially perceptive modular wind resource sensing method applicable to offshore wind farms according to the present invention.

[0021] Figure 2This is a schematic diagram of the three-dimensional wind field distribution;

[0022] Figure 3 Background image showing wind field vectors and wind speeds at equal values;

[0023] Figure 4 Configure the topology diagram for the nodes;

[0024] Figure 5 Flowchart for site selection and layout optimization;

[0025] Figure 6 A flowchart for multi-source data acquisition and preprocessing;

[0026] Figure 7 This is a diagram of the lightweight Transformer prediction model structure.

[0027] Figure 8 This is a diagram of the two-layer LSTM prediction model structure;

[0028] Figure 9 Flowchart of wind resource forecasting for the central platform. Detailed Implementation

[0029] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0030] This invention provides a modular wind resource forecasting system and its layout structure, comprising several modular wind resource acquisition devices, a central processing platform, and a deployment and communication structure supporting rapid connectivity. The modular wind resource acquisition devices are installed at multiple key structural locations within the offshore wind farm, forming a multi-point data sensing network covering the entire farm. The central processing platform is responsible for data aggregation, fusion, and forecasting. Optional relay units are used for signal enhancement in complex communication paths. The system is suitable for installation on wind turbine towers, jacket structures, or floating platforms in offshore wind farms, enabling multi-point distributed sensing and forecasting of wind speed, wind direction, and meteorological information.

[0031] The modular wind resource acquisition device includes sensor components, edge computing unit, communication module, quick-installation structure, and power supply components.

[0032] The sensor components include a three-dimensional ultrasonic anemometer, a humidity, temperature and pressure module, a radiation / visibility module, an IMU attitude module, and a power and status monitoring module, all integrated into a single protective housing.

[0033] The edge computing unit uses a low-power ARM Cortex series processor and has built-in filtering, feature extraction, sliding window construction, anomaly detection and data compression algorithms, enabling local preprocessing.

[0034] The communication module supports communication methods such as LoRa, Mesh networking, or 4G / 5G, and can automatically join the network and transmit data back to the central platform according to the installation location.

[0035] Quick-installation structure: The bottom of the integrated protective shell is equipped with a magnetic base, dovetail slot, flange fixing holes and other structural combinations, which can be quickly mounted and dismounted on different base platform surfaces.

[0036] Power supply components: Includes an integrated module of solar panel and lithium battery pack, with automatic sleep and power management functions, suitable for 24 / 7 unattended operation.

[0037] Multiple modular wind resource acquisition devices upload data to the central processing platform via a standardized communication protocol. The central processing platform first performs time alignment and spatial location calibration on the data from each node, then performs wind speed vector rotation correction and weighted fusion to form a continuous description of the wind field's spatial state. After constructing the temporal input tensor, it is fed into a trained prediction model (such as a two-layer LSTM or a lightweight Transformer) to output a wind speed trend prediction sequence for the next 1 to 6 hours.

[0038] In addition, the central processing platform dynamically evaluates model performance based on historical errors. If the continuous prediction error exceeds the threshold (e.g., MAPE>15%), it triggers online model fine-tuning or remotely sends update instructions to edge nodes to maintain the overall system's prediction accuracy and adaptability.

[0039] Please see Figure 1 The present invention also provides a spatially perceptive modular wind resource sensing method applicable to offshore wind farms, comprising:

[0040] Step 1: Deploy a three-dimensional wind field sensing system

[0041] The three-dimensional sensing deployment strategy for wind farms revolves around five types of locations that can be attached to and expanded: the middle and top sections of wind turbine towers, the edges of jacket foundation platforms, the top of substations / roofs of control centers, submarine cable outlets and edge channel platforms, and buoy arrays in the waters ahead of the prevailing wind direction. Following the three-dimensional optimization principle of "structural topology adaptability—prevailing wind direction coupling—wind power density driving," density function optimization and integer programming solutions are used to select candidate points. Under engineering constraints such as minimum point spacing, an actual deployment map is generated. A modular, plug-and-play hardware form and redundant power supply / communication design are adopted to build a three-dimensional heterogeneous sensing network that covers the entire field, is resistant to sea state disturbances, and is maintainable and expandable. This includes:

[0042] Step 1.1, node selection, including:

[0043] Based on the wind farm layout, foundation type, and wind conditions, a structural attachment + floating expansion strategy is adopted, such as... Figure 2 As shown, this invention constructs a three-dimensional wind field distribution structure covering the entire offshore wind farm area, and the following five types of points are preferred for the deployment of data acquisition modules.

[0044] 1) Middle and top sections of the wind turbine tower: Installed at a height of approximately 30–60 meters from the base, the vertical wind speed gradient and shear characteristics were collected.

[0045] 2) Edge of the jacket foundation platform: symmetrically arranged at the four corners, with local vortex structures and wind field interference.

[0046] 3) Top of the booster station / roof of the central control center: serving as a steady-state reference point for the entire site.

[0047] 4) Submarine cable exit platform and edge channel platform: to obtain low-level air disturbances.

[0048] 5) Buoy array deployment: Temporary / long-term anchored buoys are deployed in the prevailing wind direction area to form a sensing front.

[0049] Step 1.2: Based on the following four optimization principles, spatial site selection and three-dimensional layout optimization of wind turbine deployment nodes are carried out to achieve a comprehensive balance between resource utilization efficiency, structural installability, and operational safety, including:

[0050] (1) Structural topology adaptability principle: Based on the geometric distribution and installability of existing structures such as wind turbine towers, jacket platforms, substation roofs, and buoy structures within the wind farm, a structural adaptability function is constructed. It is constructed from the structural topology diagram, with deployable areas assigned a value of 1 and non-deployable areas assigned a value of 0. Intermediate values ​​are approximated by the platform surface shape function.

[0051] (2) Dominant wind direction coupling principle: Analyze the dominant wind direction of the offshore wind field through meteorological observations or historical data. ,like Figure 3 As shown, the prevailing wind direction is superimposed on the wind speed isopleth map as a wind vector. Different colored areas represent changes in the wind speed gradient. The angle between the line connecting the candidate points and the prevailing wind direction is calculated to obtain the prevailing wind direction response function.

[0052]

[0053] In the formula, r represents the spatial position vector of any point within the offshore wind farm, r0 is the spatial vector of the reference position, and r0 is the wind vector corresponding to the prevailing wind direction. Let it be its unit vector. This indicates the angle between the prevailing wind direction and the line connecting them.

[0054] (3) Wind power density driving principle: Construct wind power density P based on historical wind speeds (r) To identify areas with high wind energy resources and areas of drastic fluctuation:

[0055]

[0056] In the formula, P (r) This represents the theoretical power of wind energy swept across the area A by the wind turbine per unit time at position r; represents air density; v represents the average wind speed at location r.

[0057] (4) Comprehensive suitability scoring principle: All locations are jointly optimized using simulated wind field heat maps, terrain shading maps, and historical wind power density data to calculate the deployment density function D(r):

[0058]

[0059] In the formula, max represents taking the maximum value. This represents the weighting function for the effects of terrain and shading, reflecting the degree to which local terrain weakens the wind flow field. This represents the dominant wind direction coupled response function; α1 represents wind power density; α2, α3 represent weighting coefficients, α1+α2+α3=1, which are determined according to the actual optimization objectives (such as power generation efficiency, construction conditions or safety margin).

[0060] The system assigns priority scores to all candidate deployment nodes using the deployment density function D(r) and constructs a deployment optimization objective function, aiming to maximize global deployment efficiency under constraints. N optimal deployment nodes are selected from the set of all candidate deployment nodes, denoted as . r1 represents the first candidate deployment node, r2 represents the second candidate deployment node, M is the total number of candidate deployment nodes, and the deployment node selection variable is... , indicating whether to select the i-th point, using the following objective function:

[0061]

[0062] In the formula, D(r) i ) represents the overall suitability function, N represents the optimal number of nodes to be selected, and d min This indicates the minimum safe distance between two points.

[0063] Using integer linear programming algorithms, such as Figure 4As shown, the node layout topology forms a multi-point network covering the main structures and boundary areas of the wind farm. Under the constraint that "the distance between any two nodes is not less than a preset minimum distance threshold," N optimal nodes are selected. The specific output format is, for example, "top of tower No. 17, west mast of southern floating platform A, middle of the substation roof," forming the actual engineering layout diagram. Figure 5 As shown, the layout optimization process includes four steps: candidate point selection, density function calculation, integer programming solution, and final node placement, forming a visualized layout optimization process. The minimum spacing threshold is set according to equipment size and construction safety requirements, preferably 10m to 50m, and more preferably 15m to 30m.

[0064] Step 2: Perform multi-source data acquisition and edge preprocessing.

[0065] Multi-source data acquisition and edge preprocessing are implemented. Each deployment node is equipped with a combination of multiple sensors, including anemometers, humidity, temperature, and pressure modules, radiation / visibility modules, and IMU attitude and power status monitoring. Lightweight edge computing units are used to synchronously acquire multimodal data, perform time broadcasting and time stamp mapping alignment, and execute median filtering and first-order difference anomaly removal, sliding window convolution compression, wavelet energy-based multi-scale feature extraction, and K-means perturbation grading. The system combines streaming uplink of "main features + anomaly records + original summaries" with local cache backup. Furthermore, low-power strategies in working / sleep states, solar power + battery pack power supply, and IP65 protection structures ensure continuous and stable operation of the nodes under bandwidth constraints and harsh sea conditions.

[0066] like Figure 6 As shown, the system performs multi-source data acquisition and preprocessing at the edge nodes, including four core steps: signal acquisition, anomaly removal, feature extraction, and uplink transmission, to achieve closed-loop processing of data from acquisition to cloud fusion.

[0067] Step 2.1, the structure of the acquisition module and the configuration of the sensors, including:

[0068] Each deployment node is equipped with the following composite sensors and embedded acquisition devices, depending on its functional role and deployment environment. Among them, the anemometer is the main sensing source, and temperature, humidity, pressure and attitude information are used to correct for the influence of boundary layer height and buoy angle drift on wind speed measurement:

[0069] Table 1 Sensor Configuration

[0070] Step 2.2, Edge Device Platform Construction and Task Allocation, includes:

[0071] Each node is equipped with a lightweight edge computing unit (such as Jetson Nano, Raspberry Pi 4B, ESP32-S3 low-power MCU, etc.) and embeds modules that run the following tasks:

[0072] (1) Signal acquisition and buffering

[0073] Data from various sensors is acquired via serial port / I2C / analog channel, stored in a local memory queue according to timestamps, and the buffer size is dynamically adjusted.

[0074] (2) Anomaly removal and redundancy reduction

[0075] Median filtering and first-order difference methods were used to remove abrupt transitions; sliding window convolution compression was applied to the time series.

[0076]

[0077] In the formula, x filt (t) represents the smoothing result at time t, where t is the time index. Let w represent the data of the i-th sampling point, and w be the adjustable window width threshold.

[0078] (3) Local feature extraction and preliminary clustering

[0079] Wavelet analysis was performed on the wind speed sequence to extract multi-scale disturbance energy indices.

[0080]

[0081] In the formula, E j (t) represents the local energy at the j-th scale, and the wavelet coefficients at scale j and time i after wavelet decomposition. This represents the instantaneous energy density at scale j.

[0082] At the same time, K-means is used to classify the degree of perturbation of time-period features.

[0083] (4) Uplink preprocessing and cache backup

[0084] The processed data is divided into "main feature value + anomaly record + original summary" and uploaded to the host computer or monitoring center via LoRa or 4G at a set frequency. The original sequence is stored in CSV format on an SD card for later use.

[0085] Step 2.3, data synchronization mechanism and time alignment strategy, including:

[0086] The wind turbine deployment nodes have built-in clock modules. The wind resource sensing system synchronizes its time hourly via broadcast from the main control timer at the booster station, with a maximum error controlled within ±2 seconds. The following strategy is used for aligning the timestamps of the collected data:

[0087] Let the time deviation of each node be . Then, the uploaded data will be mapped uniformly as follows:

[0088]

[0089] In the formula, Let i be the calibrated time of the i-th node. This is the original timestamp. This represents the node clock offset, measured by the synchronization algorithm.

[0090] This mechanism ensures that multi-node data can be jointly reconstructed in the cloud in terms of space and time. After time alignment, the multi-point data is fused with a unified time reference to construct a 4D tensor form of the wind field.

[0091]

[0092] In the formula, This represents the wind field in spatial coordinates. The overall state tensor over time t, where u, v, and w are the components of wind speed in the three directions, and T, H, and P are the temperature field, humidity field, and air pressure field, respectively; the subscript t indicates the discrete time frame index.

[0093] Step 2.4, Engineering Energy Management and Equipment Endurance Guarantee, includes:

[0094] The system employs a low-power, high-frequency switching mode, divided into active and sleep states. The energy management strategy is as follows:

[0095] The power consumption of MCU-type edge devices is controlled within the range of 0.2W to 0.5W (preferably 0.3W), and it has an automatic frequency adjustment mechanism based on data fluctuations.

[0096] Each node is equipped with: a 20W solar panel, a 50Wh lithium battery pack, and a 10A charge / discharge controller, and must be able to maintain normal operation for 5-7 days under continuous cloudy and rainy weather.

[0097] The equipment structure adopts an IP65 sealed design, and the protection level is suitable for humid marine environments; redundant backup nodes are set up for key nodes (such as buoy base stations and booster station roofs) to ensure stable system operation.

[0098] Step 3: Perform spatiotemporal feature modeling and intelligent prediction.

[0099] The spatiotemporal feature modeling and intelligent prediction process involves a central platform that performs temporal alignment and spatial reconstruction of multi-node data. A weighted nearest neighbor / Gaussian kernel is used for spatial interpolation to a regular grid. Mean-standard deviation normalization and detrending processing are combined to construct a full-field spatiotemporal tensor input. In both stable and drastically disturbed regions, a dual-layer stacked LSTM and a lightweight Transformer are used for multi-step prediction of wind speed / direction and inflow conditions. Physical priors such as wind energy density and terrain shading factors are introduced for post-processing fusion correction. Simultaneously, online incremental fine-tuning triggered by a rolling MAPE threshold and a remote update mechanism for edge nodes maintain the model's adaptability and robustness in non-stationary wind conditions and structural change scenarios, including:

[0100] Step 3.1, Dataset Construction and Tensor Organization, includes:

[0101] All data reported by the optimally deployed nodes are time-aligned and spatially reconstructed by the central platform to construct the spatiotemporal tensor of the entire wind field. The specific steps are as follows:

[0102] (1) Timing window construction

[0103] Given a time step Δt, with a sliding window length T win Constructing a time-slice sequence X with step size ΔT k :

[0104]

[0105] In the formula, X k Let t represent the set of wind field tensor sequences within the k-th time window. k This indicates the end time of the k-th time window, where the subscript i is the time step index within the window, and k is the current sliding window number.

[0106] (2) Spatial interpolation and grid unification

[0107] The data of each optimally positioned node are interpolated to a regular 3D grid, using either weighted nearest neighbor interpolation or a Gaussian kernel method.

[0108]

[0109] in, Represents the reconstructed wind field tensor under a regular grid; This represents the observation data of the i-th node; N is the number of valid observation nodes; For interpolation weights, satisfying .

[0110] (3) Data normalization and detrending

[0111] To improve training stability, mean-standard deviation normalization is used:

[0112]

[0113] in, This represents the normalized signal; These are the original time series values; This represents the sample mean; This represents the sample standard deviation and is used to fit a trend term to historical daily periodic data. It is used to eliminate seasonal fluctuations.

[0114] Step 3.2, Prediction Model Structure Design, including:

[0115] The central platform integrates the following two types of prediction models, combining the spatiotemporal structure of the wind field with historical evolution patterns, to perform multi-step predictions of wind speed and direction:

[0116] (1) LSTM network based on double-layer recurrent structure

[0117] like Figure 8 As shown, the dual-layer stacked LSTM model structure includes an input layer, two recurrent layers, and an output mapping layer, which is used for short-term trend prediction of wind speed sequences in stable regions.

[0118] The input to this LSTM network based on a two-layer recurrent structure is a normalized multidimensional time series tensor. .

[0119] The structure of this LSTM network based on a two-layer recurrent structure is as follows: It consists of two stacked LSTM layers, suitable for long-term, stable wind conditions, with fast training speed and a small number of parameters. The hidden dimension of this LSTM network based on the two-layer recurrent structure is H=128, and it outputs a wind speed prediction sequence.

[0120]

[0121] in, This represents the predicted future T-step wind speed sequence; represents the normalized multidimensional input tensor, which includes features such as wind speed, temperature, and humidity; LSTM1 and LSTM2 represent the first and second long short-term memory units, respectively; t is the current time index; and T is the prediction step size.

[0122] Table 2 Parameters of the Two-Layer LSTM Model

[0123] (2) Lightweight Transformer Network

[0124] like Figure 7As shown, the lightweight Transformer network employs temporal location encoding, multi-head attention, and a feedforward network structure, which can capture long-term dependent wind speed features while maintaining computational lightweightness.

[0125] Input encoder: Employs a concatenation of temporal position encoding (PE(t)) and multidimensional feature encoding.

[0126]

[0127] In the formula, E in The input matrix of the Transformer network is composed of temporal location encoding and feature vectors. Represents the normalized feature tensor of the current time; PE(t) is the time position code used to preserve temporal position information.

[0128] Core structure: 3-layer attention module, using causal masking to control the direction of information flow;

[0129] Output module: Linear mapping outputs future T pred Wind speed and direction at each time step:

[0130]

[0131] In the formula, Transformer(·) represents the main function of the self-attention network, which contains 3 attention modules and a feedforward layer.

[0132] Table 3 Parameters of Lightweight Transformer Model

[0133] (3) Model selection strategy:

[0134] Lightweight Transformer networks are used in areas with severe fluctuations, such as buoys / jacket structures; LSTM networks based on a two-layer loop structure are used in stable areas, such as towers / booster stations.

[0135] Step 3.3, Physical Enhancement Modeling and Prediction Correction Mechanism, includes:

[0136] To improve the generalization ability of the prediction model under non-stationary wind conditions, a wind power physical driving factor and a terrain influence function are introduced as priors:

[0137] (1) Wind energy density driver:

[0138]

[0139] In the formula, P wind (t) represents the wind energy power density at time t; Where A is the air density; A is the swept area of ​​the wind turbine. The average wind speed is used as an auxiliary feature input to an LSTM network with a two-layer recurrent structure or a lightweight Transformer network to guide the model to focus on the nonlinear characteristics of the wind speed's impact on energy.

[0140] (2) Terrain shading function (shading factor):

[0141]

[0142] in, S is the terrain shading correction factor. obst The projected area of ​​the obstacle is used to correct the wind speed prediction bias; S total This indicates the total sampling area of ​​the region.

[0143] (3) Fusion correction formula (post-processing):

[0144]

[0145] in, This represents the corrected final wind speed forecast. This indicates the model's predicted wind speed (from an LSTM network model or a Transformer network model). This is a dynamic compensation term based on local historical errors.

[0146] Step 3.4, online update and model adaptation mechanism, including:

[0147] like Figure 9 As shown, the wind resource prediction process of the central platform includes four stages: data acquisition, spatiotemporal modeling, predictive inference and model adaptive update, forming a complete closed loop from perception to feedback.

[0148] The system incorporates online evaluation and model self-learning mechanisms during operation, including the following three core modules:

[0149] (1) Error monitoring module: Calculates the rolling MAPE index:

[0150]

[0151] In the formula, MAPE k Indicates the first The moving average absolute percentage error over each monitoring period; This represents the number of samples within the current monitoring period. This indicates the model's predicted wind speed; This indicates the measured wind speed.

[0152] (2) Triggering mechanism: If 3 consecutive cycles , To preset the error threshold, If the range is 10%-20%, preferably 15%, then the following operation is performed:

[0153] (3) Online incremental fine-tuning (using the most recent data window): Local updates of the weights of the wind resource prediction model based on the latest data samples.

[0154] Remotely distribute wind resource prediction model parameters: The main control center sends the corrected model weights to each optimal deployment node.

[0155] Abnormal node restart and recalibration commands: For deployed nodes with significantly excessive errors, resampling and time synchronization commands are automatically issued to restore synchronization performance.

[0156] Step 4: Conduct wind energy assessment and system feedback optimization.

[0157] The wind energy assessment and system feedback optimization module converts the predicted wind field into the regional total wind energy density per unit area and three-dimensional integral, generating a wind energy heat map, trend curve, and power output prediction. Based on the power curve / operational efficiency parameters, it implements short-term scheduling, energy storage charging / discharging, and grid connection curve correction. Simultaneously, it constructs a "prediction-measurement" closed loop using weighted relative error (WRE), triggering online model updates, point-of-sight adjustments, and regional expansion / reduction suggestions. Through dynamic evaluation of information gain, signal stability, and deployment structure values, it periodically optimizes the sensing network by adding or deleting elements, ensuring a self-consistent closed loop of "sensing-prediction-assessment-feedback" and continuously improving the accuracy of wind field perception and operational benefits, including:

[0158] Step 4.1, wind energy resource distribution assessment, including:

[0159] Based on the predicted wind speed sequences of each optimal deployment node Calculate the local wind energy density distribution to generate a wind farm heat map, which can be used for spatial assessment and power dispatching reference.

[0160]

[0161] in, air density; Represents spatial coordinates Wind speed sequence over time t; This refers to the wind energy density per unit area.

[0162] The total instantaneous wind energy P can be estimated by performing three-dimensional integration over the entire deployment area. total (t):

[0163]

[0164] The wind resource sensing system generates the following: regional wind energy heat map (visual analysis of hot and weak areas); time-series wind energy trend map (analysis of diurnal and extreme changes); and wind energy output prediction curve (providing an interface for grid-side power curves).

[0165] Step 4.2, prediction accuracy closed-loop feedback mechanism, including:

[0166] The wind resource sensing system constructs a closed-loop feedback channel based on prediction errors and actual measurement errors, enabling the wind resource prediction network model to adapt and the deployment strategy to fine-tune. The core indicator is the weighted relative error (WRE).

[0167]

[0168] Among them, WRE i This represents the weighted relative error index; w i Energy weighting factor for location (based on wind power density contribution):

[0169]

[0170] When the average WRE within a continuous window is greater than a preset error threshold , The typical value range is 5%-20%, with 10% being preferred. When the volatility is high, the wind resource prediction system will trigger the following feedback strategies: online incremental model update (see step 3.4); point perception adjustment suggestions: such as "the signal of buoy ST-03 fluctuates greatly, it is recommended to replace the sensor or adjust the height to +5m"; regional structure optimization instructions: such as "the wind energy in the T-12 area of ​​the tower changes drastically, it is recommended to increase the number of boundary buoy points".

[0171] Step 4.3, Wind power prediction and dispatch optimization:

[0172] The wind speed forecast is converted into a wind turbine output estimate. Considering the wind turbine power curve P(u), wind power prediction is performed:

[0173]

[0174] in, This represents the predicted total output power at time t. This represents the power value corresponding to the power curve of the i-th wind turbine; denoted as the operating efficiency parameter of the i-th wind turbine; N represents the total number of wind turbines; the prediction results are used to guide short-term wind energy dispatch, charging and discharging energy storage strategies, and grid connection curve correction.

[0175] Step 4.4, multi-point dynamic adjustment mechanism, including:

[0176] The wind resource forecasting system supports a periodic optimization and addition / removal mechanism for deployed nodes, which utilizes a dynamic evaluation function. A comprehensive evaluation of the operational status and contribution of each deployed node is conducted to achieve adaptive optimization and adjustment of the wind field sensing network.

[0177]

[0178] in, Indicates the first The overall performance score of each deployed node. This represents the information gain (reduction in prediction error) of the i-th deployed node. This represents the signal stability index (volatility, packet loss rate) of the i-th deployment node. This represents the layout structure value of the i-th deployment node (refer to the topology function description in step 1.2). α, β, and γ represent weight coefficients, which are set according to the system optimization objective.

[0179] Wind resource forecasting system based on Dynamic trend analysis and automatic identification of efficient and inefficient nodes: When the node performance score is lower than the preset threshold θ, the wind resource prediction system automatically marks "inefficient points" and provides suggestions for layout adjustment, node status adjustment, and new node optimization to maintain the optimal state of the overall wind field perception capability of the wind resource prediction system; newly added points are reselected according to the node deployment optimization function in step 1 to keep the overall wind field perception capability optimal.

[0180] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A spatially perceptive modular wind resource sensing method suitable for offshore wind farms, characterized in that, include: Step 1, Deploy a three-dimensional wind field sensing system, including: Step 1.1, Node selection; Step 1.2: Perform spatial site selection and three-dimensional layout optimization for wind turbine deployment nodes; Step 2 involves multi-source data acquisition and edge preprocessing, including: Step 2.1: Configure multiple sensors; Step 2.2, Edge device platform construction and task allocation; Step 2.3: Construct a data synchronization mechanism and time alignment strategy; Step 3, perform spatiotemporal feature modeling and intelligent prediction, including: Step 3.1: Construct the dataset and organize tensors; Step 3.2: Design the prediction model structure; Step 3.3: Perform physical enhancement modeling and construct a prediction correction mechanism; Step 3.4: Construct an online update and model adaptation mechanism; Step 4, conduct wind energy assessment and system feedback optimization, including: Step 4.1: Conduct a wind energy resource distribution assessment; Step 4.2: Construct a closed-loop feedback mechanism for prediction accuracy; Step 4.3: Perform wind power prediction and scheduling optimization; Step 4.4: Construct a multi-point dynamic adjustment mechanism.

2. The spatially perceptive modular wind resource sensing method applicable to offshore wind farms according to claim 1, characterized in that, Step 1.1 includes: Deploy data collection modules for the following five types of locations: a. Middle and top sections of the wind turbine tower; b. Edge of the jacket foundation platform; c. Top of the booster station / roof of the control center; d. Submarine cable outgoing platform and edge channel platform; e. Buoy array deployment.

3. The spatially perceptive modular wind resource sensing method applicable to offshore wind farms according to claim 1, characterized in that, Step 1.2 includes: The spatial location and three-dimensional layout of wind turbine deployment nodes are optimized based on the following four optimization principles: (1) Structural topology adaptability principle: Based on the geometric distribution and installability of wind turbine towers, jacket platforms, substation roofs, and buoy structures within the wind farm, a structural adaptability function is constructed. It is constructed from the structural topology diagram, with deployable areas assigned a value of 1 and non-deployable areas assigned a value of 0. Intermediate values ​​are approximated by the platform surface shape function. (2) Dominant wind direction coupling principle: Analyze the dominant wind direction of the offshore wind field. Calculate the angle between the line connecting the candidate points and the target point to obtain the prevailing wind direction response function. : In the formula, r represents the spatial position vector of any point within the offshore wind farm, r0 is the spatial vector of the reference position, and r0 is the wind vector corresponding to the prevailing wind direction. Let it be its unit vector. Indicates the angle between the prevailing wind direction and the line connecting them; (3) Wind power density driving principle: Construct wind power density P based on historical wind speeds (r) To identify areas with high wind energy resources and areas of drastic fluctuation: In the formula, The air density is represented by 'r'; the average wind speed at location 'r' is represented by 'v'. (4) Comprehensive suitability scoring principle: All locations are jointly optimized using simulated wind field heat maps, terrain shading maps, and historical wind power density data to calculate the deployment density function D(r): In the formula, max represents taking the maximum value, and w s (r) represents the weighting function for the influence of terrain and shading, w w (r) represents the dominant wind direction coupled response function; α1, α2, and α3 represent weighting coefficients; All candidate deployment nodes are assigned priority scores using the deployment density function D(r), and a deployment optimization objective function is constructed to maximize global deployment efficiency under constraints. N optimal deployment nodes are selected from the set of all candidate deployment nodes, denoted as . r1 represents the first candidate deployment node, r2 represents the second candidate deployment node, M is the total number of candidate deployment nodes, and the deployment node selection variable is... , indicating whether to select the first For each point, the following objective function is used: In the formula, D(r) i ) represents the overall suitability function, N represents the optimal number of nodes to be selected, and d min This indicates the minimum safe distance between two points; Using an integer linear programming algorithm, N optimal nodes are selected under the constraint that "the distance between any two nodes is not less than the preset minimum distance threshold", thus forming the actual engineering layout diagram.

4. The spatially perceptive modular wind resource sensing method applicable to offshore wind farms according to claim 1, characterized in that, In step 2.1, the following sensors are configured: a three-dimensional ultrasonic anemometer, a humidity, temperature and pressure module, a radiation / visibility module, an IMU attitude module, and a power supply and status monitoring module. Step 2.2 includes: Each node is equipped with a lightweight edge computing unit, embedding the following task modules: (1) Signal acquisition and buffering Data from various sensors is acquired via serial port / I2C / analog channel and stored in a local memory queue according to timestamps, with the buffer size dynamically adjusted. (2) Anomaly removal and redundancy reduction Median filtering and first-order difference methods are used to remove abrupt transitions, and sliding window convolution compression is applied to the time series. In the formula, x filt (t) represents the smoothing result at time t, where t is the time index. This represents the data of the i-th sampling point, where w is the adjustable window width threshold; (3) Local feature extraction and preliminary clustering Wavelet analysis was performed on the wind speed sequence to extract multi-scale disturbance energy indices. In the formula, E j (t) represents the local energy at the j-th scale, and the wavelet coefficients at scale j and time i after wavelet decomposition. This represents the instantaneous energy density at the j-th scale layer number; At the same time, K-means is used to classify the degree of perturbation of time-period features; (4) Uplink preprocessing and cache backup The processed data is divided into "main feature value + anomaly record + original summary" and uploaded to the host computer or monitoring center via wireless communication module at a set frequency.

5. A spatially perceptive modular wind resource sensing method suitable for offshore wind farms according to claim 1, characterized in that, Step 2.3 includes: The wind turbine deployment nodes have built-in clock modules. The wind resource sensing system synchronizes its time according to a set cycle via broadcast time from the main control time of the booster station. The maximum error is controlled within a set threshold range based on project requirements. The following strategy is used for aligning the timestamps of the collected data: Let the time deviation of each node be . The uploaded data will be mapped uniformly as follows: In the formula, For the first The calibrated time of each node This is the original timestamp. This refers to the node clock offset; Multi-point data, after time alignment, are fused using a unified time base to construct a 4D tensor of the wind field: In the formula, This represents the wind field in spatial coordinates. The overall state tensor over time t, where u, v, and w are the components of wind speed in the three directions, and T, H, and P are the temperature field, humidity field, and air pressure field, respectively; the subscript t indicates the discrete time frame index.

6. The spatially perceptive modular wind resource sensing method applicable to offshore wind farms according to claim 1, characterized in that, Step 3.1 includes: All data reported by the optimally deployed nodes are time-aligned and spatially reconstructed by the central platform to construct the spatiotemporal tensor of the entire wind field. The specific steps are as follows: Step 3.1.1, Timing Window Construction Given a time step Δt, with a sliding window length T win Constructing a time-slice sequence X with step size ΔT k : In the formula, X k Let t represent the set of wind field tensor sequences within the k-th time window. k This indicates the end time of the k-th time window, where the subscript i is the time step index within the window, and k is the current sliding window number; Step 3.1.2, Spatial Interpolation and Mesh Unification The data of each optimally positioned node are interpolated to a regular 3D grid, using either weighted nearest neighbor interpolation or a Gaussian kernel method. in, Represents the reconstructed wind field tensor under a regular grid; This represents the observation data of the i-th node; N is the number of valid observation nodes; For interpolation weights, satisfying ; Step 3.1.3, Data Normalization and Detrending Normalization using mean-standard deviation: in, This represents the normalized signal; These are the original time series values; This represents the sample mean; This represents the sample standard deviation and is used to fit a trend term to historical daily periodic data. .

7. A spatially perceptive modular wind resource sensing method applicable to offshore wind farms according to claim 1, characterized in that, Step 3.2 includes: The central platform integrates the following two types of prediction models, combining the spatiotemporal structure of the wind field with historical evolution patterns, to perform multi-step predictions of wind speed and direction: (1) LSTM network based on double-layer recurrent structure The input to this LSTM network based on a two-layer recurrent structure is a normalized multidimensional time series tensor. ; The structure of this LSTM network based on a two-layer recurrent structure includes: two stacked LSTM layers; The hidden dimension of this LSTM network based on a two-layer recurrent structure is H=128, and it outputs a wind speed prediction sequence: ， in, This represents the predicted future T-step wind speed sequence; This represents the normalized multidimensional input tensor, containing features such as wind speed, temperature, and humidity; LSTM1 and LSTM2 represent the first and second long short-term memory units, respectively; t is the current time index; and T is the prediction step size. (2) Lightweight Transformer Network The input encoder of this lightweight Transformer network uses a concatenation of temporal position encoding PE(t) and multidimensional feature encoding. In the formula, E in The input matrix of the Transformer network is composed of temporal location encoding and feature vectors. Represents the normalized feature tensor of the current time step; PE(t) is the temporal position code used to preserve temporal position information; The core structure of this lightweight Transformer network consists of three attention modules that use causal masks to control the direction of information flow. The output module of this lightweight Transformer network: a linear mapping outputs the future T. pred Wind speed and direction at each time step: ， In the formula, Transformer(·) represents the main function of the self-attention network, which includes 3 attention modules and a feedforward layer; (3) Model selection strategy: A lightweight Transformer network is used for regions with drastic fluctuations; an LSTM network based on a two-layer recurrent structure is used for regions with stable conditions.

8. A spatially perceptive modular wind resource sensing method applicable to offshore wind farms according to claim 1, characterized in that, Step 3.3 includes: Introduce the physical driving factor of wind power and the topographic influence function as priors: (1) Wind energy density driver: In the formula, P wind (t) represents the wind energy power density at time t; Where A is the air density; A is the swept area of ​​the wind turbine. Average wind speed; As auxiliary feature inputs, LSTM networks or lightweight Transformer networks with a two-layer recurrent structure are used to guide the model network to focus on the nonlinear characteristics of the influence of wind speed on energy. (2) Terrain shading function: in, S is the terrain shading correction factor. obst The projected area of ​​the obstacle is used to correct the wind speed prediction bias; S total This represents the total sampling area of ​​the region; (3) Fusion correction formula: in, This represents the corrected final wind speed forecast. This indicates that the model predicts wind speed. This is a dynamic compensation term based on local historical errors; Step 3.4 includes: The system establishes an online evaluation and model self-learning mechanism, comprising the following three core modules: (1) Error monitoring module: Calculates the rolling MAPE index: In the formula, MAPE k This represents the moving average absolute percentage error for the k-th monitoring period; T is the number of samples in the current monitoring period. This indicates the model's predicted wind speed; This indicates the measured wind speed; (2) Triggering mechanism: When the model error exceeds the preset error threshold for multiple consecutive periods. Automatically triggers model adaptive update operation: (3) Adaptive update operation: Online incremental fine-tuning: Locally updating the weights of the wind resource prediction model based on the latest data samples; Remotely distribute wind resource prediction model parameters: The main control center sends the corrected model weights to each optimal deployment node; Abnormal node restart and recalibration commands: For deployed nodes with significantly excessive errors, resampling and time synchronization commands are automatically issued to restore synchronization performance.

9. A spatially perceptive modular wind resource sensing method applicable to offshore wind farms according to claim 1, characterized in that, Step 4.1 includes: Based on the predicted wind speed sequences of each optimal deployment node Calculate the local wind energy density distribution to generate a wind farm heat map, which can be used for spatial assessment and power dispatching reference. in, air density; Represents spatial coordinates Wind speed sequence over time t; Wind energy density per unit area; The total instantaneous wind energy P can be estimated by performing three-dimensional integration over the entire deployment area. total (t): The wind resource sensing system generates the following based on this: regional wind energy heat map, time-series wind energy trend map, and wind energy output prediction curve. Step 4.2 includes: The wind resource sensing system constructs a closed-loop feedback channel based on prediction errors and actual measurement errors to achieve adaptive fine-tuning of the wind resource prediction network model. The core indicator is the weighted relative error. Among them, WRE i This represents the weighted relative error index; w i Point energy weighting factor: When the average WRE within a continuous window exceeds the preset error threshold At that time, the wind resource forecasting system triggers the following feedback strategy: The weights of the wind speed prediction model are incrementally updated based on the latest observation data to achieve self-learning and accuracy correction during operation; when the error of a specific measuring point continues to exceed the limit, parameter adjustment suggestions are automatically generated; when the wind energy in a local area changes drastically or the boundary energy gradient is significant, the wind resource prediction system generates deployment optimization suggestions.

10. A spatially perceptive modular wind resource sensing method applicable to offshore wind farms according to claim 1, characterized in that, Step 4.3 includes: The wind speed forecast is converted into a wind turbine output estimate, taking into account the wind turbine power curve P(u), to calculate the wind power output. predict: in, This represents the predicted total output power at time t. This represents the power value corresponding to the power curve of the i-th wind turbine; represents the operating efficiency parameter of the i-th wind turbine; N is the total number of wind turbines; the prediction results are used to guide short-term wind energy dispatch, charging and discharging energy storage strategies, and grid connection curve correction. Step 4.4 includes: The wind resource forecasting system supports a periodic optimization and addition / removal mechanism for deployed nodes, which utilizes a dynamic evaluation function. A comprehensive evaluation of the operational status and contribution of each deployed node is conducted to achieve adaptive optimization and adjustment of the wind field sensing network. in, This represents the overall performance score of the i-th deployed node. This represents the information gain of the i-th deployment node; This represents the signal stability index of the i-th deployment node; Indicates the first The layout structure value of each deployment node; α, β, and γ represent weight coefficients, which are set according to the system optimization objective; Wind resource forecasting system based on Dynamic trend analysis and automatic identification of efficient and inefficient nodes: When the node performance score is lower than the preset threshold θ, the wind resource prediction system automatically marks "inefficient points" and provides suggestions for layout adjustment, node status adjustment, and new node optimization to maintain the optimal state of the overall wind field perception capability of the wind resource prediction system; newly added points are reselected according to the node deployment optimization function in step 1 to keep the overall wind field perception capability at its best.

Images