A method, apparatus, equipment and medium for rapid assessment of wind resources in complex mountainous areas
By constructing a multi-channel terrain condition tensor and encoder-decoder architecture, the dependence on numerical solutions and external background wind fields in complex mountain wind resource assessment is solved, enabling fast and accurate wind resource assessment under preset inflow conditions, thus improving assessment efficiency and accuracy.
Patent Information
- Application Number
- CN202611553500.7
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-10-09
- Publication Date
- 2026-11-10
AI Technical Summary
Existing technologies rely heavily on numerical solutions and external background wind fields in the assessment of wind resources in complex mountainous areas. They lack sufficient topographical representation, struggle to balance the overall wind field pattern with local high-frequency structures, and fail to adequately integrate target height with the decoding process, resulting in low assessment efficiency and poor accuracy.
A multi-channel terrain condition tensor is constructed, and an encoder and decoder architecture is combined. Through multi-scale frequency modulation encoding and decoding, rapid evaluation without CFD initial flow field and external background wind field is achieved. By utilizing relative elevation, surface roughness and inflow channel cumulative features, terrain features are captured and injected into the target height for accurate prediction.
It significantly improves the efficiency of wind resource assessment in complex mountainous areas, enabling accurate determination of wind resources with only topographic data and target altitude input, reducing calculation cycles and the need for professional personnel.
Smart Images

Figure CN123068871A_ABST
Abstract
Description
Technical Field
[0001] This application relates to the field of wind energy resource assessment technology, and in particular to a method, apparatus, equipment and medium for rapid assessment of wind resources in complex mountainous areas. Background Technology
[0002] Rapid wind resource assessment is a prerequisite for macro-level site selection, micro-level turbine deployment, turbine load prediction, and power generation calculation for wind farms. Traditional assessments of wind resources in complex mountainous areas typically establish a computational domain based on digital elevation models, surface roughness, and boundary flows, using computational fluid dynamics or coupled micro- and meso-scale models for mesh generation and iterative solutions. However, each candidate site, wind direction, or height scheme usually requires resetting boundary conditions and performing numerical iterations; as the number of sites increases or turbine deployment schemes are frequently adjusted, the computational cycle and the investment of professional personnel increase significantly. In recent years, deep learning-based wind farm proxy or downscaling models have emerged, which learn the mapping between input conditions and the flow field, using neural network forward inference to replace some repetitive numerical solutions. However, existing technologies still have the following shortcomings for rapid screening scenarios in complex mountainous terrain: 1. Online assessment is heavily reliant on numerical solutions or external background wind fields: Traditional CFD, micro-scale coupling, and some intelligent enhancement schemes still require running numerical models during the assessment phase, or require input of low-resolution background wind fields, initial flow fields, and other external field information, making it difficult to directly complete rapid field-level reconstruction with only topographic data and preset inflow conditions. 2. Insufficient task-oriented representation of topographic conditions: Single elevation, slope, or general topographic feature vectors can describe local geometry, but they are insufficient for explicit representation of surface roughness and continuous uplift, shading, and channel effects along the inflow direction, making it difficult to distinguish terrain with similar elevation differences but different inflow path structures at the model input. 3. Difficulty in simultaneously considering global wind field patterns and local high-frequency structures: Ordinary convolution or standard attention structures tend to preferentially fit smoother large-scale main distributions, while local structures such as slope top acceleration, canyon jets, shear layers, separation and reattachment, local backflow, and vertical velocity abrupt changes are easily over-smoothed. 4. Insufficient integration of target height with the decoding process: Some solutions use the target height as a post-processing interpolation parameter or train multiple fixed-height models separately. This makes it difficult for the network to directly learn the correlation between different height slices and the differences in local terrain response, while also increasing model maintenance costs. Therefore, a rapid evaluation technique is needed that can be performed under preset inflow conditions without re-executing CFD iterations in the online phase and without requiring external coarse-resolution background wind fields. Summary of the Invention
[0003] In view of this, the purpose of this application is to provide a method, device, equipment and medium for rapid assessment of wind resources in complex mountainous areas. In the online stage, there is no need to provide the initial CFD flow field, external background wind field or perform CFD iterative solution, thereby significantly improving the assessment efficiency of wind resources in complex mountainous areas. Through the collaborative representation of multi-channel terrain conditions and target height, combined with the architecture of encoder and decoder, the assessment results of mountainous wind resources can be accurately determined under the premise of only inputting terrain data and target height.
[0004] This application provides a method for rapid assessment of wind resources in complex mountainous areas, the method comprising: Construct a multi-channel terrain condition tensor under a preset incoming flow condition; wherein, the multi-channel terrain condition tensor includes at least relative elevation, surface roughness, and cumulative positive elevation difference characteristics of the incoming flow channel; The multi-channel terrain condition tensor is input into a pre-trained wind field evaluation model. The multi-scale frequency modulation encoding of the multi-channel terrain condition tensor is performed on the encoder to obtain shared encoding features. The shared encoding features and the injected target height are decoded on the decoder to obtain the multi-channel wind field prediction results. The multi-channel wind field prediction results include pressure disturbance, turbulent kinetic energy and wind speed in three directions. The multi-channel wind field prediction results are inversely normalized to obtain a rapid assessment of wind resources in complex terrain at the target height.
[0005] In one possible implementation, a multi-channel terrain condition tensor under a preset incoming flow condition is constructed through the following steps: Obtain the digital elevation model and surface roughness data of the target area, and calculate the relative elevation based on the digital elevation model; Based on the preset incoming flow conditions and the digital elevation model, the cumulative characteristics of the incoming flow channel along the incoming flow direction are calculated; The relative elevation, surface roughness, and cumulative features of the incoming flow channels are combined to construct the multi-channel terrain condition tensor.
[0006] In one possible implementation, the step of obtaining shared coding features by multi-scale frequency modulation coding of the multi-channel terrain condition tensor based on the encoder includes: The multi-channel terrain condition tensor is converted into M×M encoded features through convolutional feature extraction and scale adaptation, while retaining both low-resolution and high-resolution terrain features. The encoded features are divided and mapped into a token sequence according to a preset patch, and position information is added; The token sequence is jointly encoded by attention main response, supplementary residuals, and frequency domain dynamic scaling to obtain shared coding features.
[0007] In one possible implementation, the joint encoding of the token sequence, which includes attention main response, supplementary residuals, and frequency domain dynamic scaling, to obtain shared coding features, includes: A linear mapping is performed on the token sequence to obtain query features, key features, and value features. An attention matrix is then calculated based on the query features and key features. A dynamic gating system determined by the current token sequence is generated, and the attention master response and supplementary residual are fused according to the dynamic gating system to obtain attention fusion features; The token sequence is transformed in the frequency domain, and the features after the frequency domain transformation are dynamically scaled and grouped according to the content-related routing weights. Then, the frequency modulation features are mapped back to the spatial domain through inverse frequency domain transformation. The token sequence, the attention fusion feature, and the frequency modulation feature are weighted and combined according to the fusion coefficient to obtain the output feature of the current coding block. Multiple coding blocks are concatenated to determine the shared coding feature.
[0008] In one possible implementation, the decoding of the shared coded features and the injected target height based on the decoder to obtain multi-channel wind field prediction results includes: The shared coding features are input into five independent decoding branches respectively. In the low-resolution stage, low-resolution terrain features are fused and a first set of height modulation parameters generated by the target height are injected. In the high-resolution stage, high-resolution terrain features are fused and a second set of height modulation parameters generated by the target height are injected. Each branch generates a prediction field. Each predicted field is restored at high resolution and aligned to the engineering grid to obtain multi-channel wind field prediction results.
[0009] In one possible implementation, each independent decoding branch generates the corresponding prediction field using a base-residual form:
[0010] in, For the category of the predicted field, , This represents the prediction result for the c-th category at the current decoding scale. The base field components represent a relatively smooth background distribution. This is a correction component formed due to the influence of local topography. These correspond to pressure disturbance, turbulent kinetic energy, wind speed components in the x, y, and z directions, respectively.
[0011] In one possible implementation, the wind field assessment model is determined through the following steps: Acquire training samples, which include multi-channel terrain condition samples, target height samples, and a five-channel physical quantity real field obtained through computational fluid dynamics; The real field of the five physical quantities is subjected to channel-independent normalization to obtain supervisory data; The multi-channel terrain condition samples and target height samples are input into the neural network model to obtain five-channel prediction results; The error between the five-channel prediction results and the supervision data is calculated using a loss function. The network parameters of the neural network model are updated uniformly based on the error, so that the shared encoder and each independent decoding branch are trained together to determine the wind field evaluation model.
[0012] This application embodiment also provides a rapid assessment device for wind resources in complex mountainous areas, the rapid assessment device for wind resources in complex mountainous areas includes: The condition construction module is used to construct a multi-channel terrain condition tensor under a preset incoming flow condition; wherein, the multi-channel terrain condition tensor includes at least relative elevation, surface roughness, and cumulative positive elevation difference features of the incoming flow channel; The wind field assessment module is used to input the multi-channel terrain condition tensor into a pre-trained wind field assessment model, perform multi-scale frequency modulation encoding on the multi-channel terrain condition tensor based on the encoder to obtain shared encoding features, and decode the shared encoding features and the injected target height based on the decoder to obtain multi-channel wind field prediction results; wherein, the multi-channel wind field prediction results include pressure disturbance, turbulent kinetic energy and wind speed in three directions; The output module is used to perform inverse normalization processing on the multi-channel wind field prediction results to obtain a rapid assessment of wind resources in complex terrain at the target height.
[0013] This application also provides an electronic device, including: a processor, a memory, and a bus. The memory stores machine-readable instructions executable by the processor. When the electronic device is running, the processor communicates with the memory via the bus. When the machine-readable instructions are executed by the processor, the steps of the complex mountain wind resource rapid assessment method described above are performed.
[0014] This application also provides a computer-readable storage medium storing a computer program that, when executed by a processor, performs the steps of the complex mountain wind resource rapid assessment method described above.
[0015] This application provides a method, apparatus, device, and medium for rapid assessment of wind resources in complex mountainous areas. The method includes: constructing a multi-channel topographic condition tensor under a preset incoming flow condition; wherein the multi-channel topographic condition tensor includes at least relative elevation, surface roughness, and cumulative positive elevation difference features of the incoming flow channel; inputting the multi-channel topographic condition tensor into a pre-trained wind field assessment model; performing multi-scale frequency modulation encoding on the multi-channel topographic condition tensor based on an encoder to obtain shared encoding features; decoding the shared encoding features and the injected target height based on a decoder to obtain a multi-channel wind field prediction result; wherein the multi-channel wind field prediction result includes pressure disturbance, turbulent kinetic energy, and wind speed in three directions; and performing inverse normalization on the multi-channel wind field prediction result to obtain a rapid assessment of wind resources in complex terrain at the target height. The online phase does not require providing the initial CFD flow field, external background wind field, or performing CFD iterative solutions, thus significantly improving the efficiency of wind resource assessment in complex mountainous areas. Through the collaborative representation of multi-channel terrain conditions and target height, combined with the encoder and decoder architecture, the assessment results of mountainous wind resources can be accurately determined with only terrain data and target height as input.
[0016] To make the above-mentioned objectives, features and advantages of this application more apparent and understandable, preferred embodiments are described below in detail with reference to the accompanying drawings. Attached Figure Description
[0017] To more clearly illustrate the technical solutions of the embodiments of this application, the accompanying drawings used in the embodiments will be briefly introduced below. It should be understood that the following drawings only show some embodiments of this application and should not be regarded as a limitation of the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.
[0018] Figure 1 A flowchart illustrating a rapid assessment method for complex mountain wind resources provided in this application embodiment; Figure 2 This is one of the structural schematic diagrams of a rapid assessment device for complex mountain wind resources provided in the embodiments of this application; Figure 3 This is a second schematic diagram of a complex mountain wind resource rapid assessment device provided in an embodiment of this application; Figure 4 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application. Detailed Implementation
[0019] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. The components of the embodiments of this application described and shown in the accompanying drawings can generally be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of this application provided in the accompanying drawings is not intended to limit the scope of the claimed application, but merely represents selected embodiments of this application. Based on the embodiments of this application, every other embodiment obtained by those skilled in the art without inventive effort falls within the scope of protection of this application.
[0020] First, the applicable scenarios for this application will be introduced. This application can be applied to the field of wind energy resource assessment technology.
[0021] Research has revealed the following technical problems in existing technologies: 1. Online evaluation is heavily reliant on numerical solutions or external background wind fields: Traditional CFD, micro-scale coupling, and some intelligent enhancement schemes still require running numerical models during the evaluation phase, or require input of low-resolution background wind fields, initial flow fields, and other external field information, making it difficult to directly complete rapid field-level reconstruction with only topographic data and preset inflow conditions. 2. Insufficient task-oriented representation of topographic conditions: Single elevation, slope, or general topographic feature vectors can describe local geometry, but they are insufficient for explicit representation of surface roughness and continuous uplift, shading, and channel effects along the inflow direction, making it difficult to distinguish topographic features with similar elevation differences but different inflow path structures at the model input. 3. Difficulty in simultaneously capturing global wind field patterns and local high-frequency structures: Ordinary convolution or standard attention structures tend to preferentially fit smoother large-scale master distributions, while local structures such as slope-top acceleration, canyon jets, shear layers, separation and reattachment, local backflow, and vertical velocity abrupt changes are easily over-smoothed. 4. Insufficient integration of target height with the decoding process: Some solutions use the target height as a post-processing interpolation parameter or train multiple fixed-height models separately. This makes it difficult for the network to directly learn the correlation between different height slices and the differences in local terrain response, while also increasing model maintenance costs. Therefore, a rapid evaluation technique is needed that can be performed under preset inflow conditions without re-executing CFD iterations in the online phase and without requiring external coarse-resolution background wind fields.
[0022] Based on this, this application provides a method for rapid assessment of wind resources in complex mountainous areas. In the online stage, there is no need to provide the initial CFD flow field, external background wind field, or perform CFD iterative solution, thereby significantly improving the assessment efficiency of wind resources in complex mountainous areas. Through the collaborative representation of multi-channel terrain conditions and target height, combined with the architecture of encoder and decoder, the assessment results of mountainous wind resources can be accurately determined under the premise of only inputting terrain data and target height.
[0023] Please see Figure 1 , Figure 1 A flowchart illustrating a rapid assessment method for complex mountain wind resources provided in an embodiment of this application. Figure 1 As shown in the embodiments of this application, the method for rapid assessment of complex mountain wind resources includes: S101: Construct a multi-channel terrain condition tensor under a preset incoming flow condition; wherein, the multi-channel terrain condition tensor includes at least relative elevation, surface roughness, and cumulative positive elevation difference features of the incoming flow channel.
[0024] In one possible implementation, a multi-channel terrain condition tensor under a preset incoming flow condition is constructed through the following steps: Obtain the digital elevation model and surface roughness data of the target area, and calculate the relative elevation based on the digital elevation model; calculate the cumulative characteristics of the incoming flow channel along the direction of the incoming flow based on the preset incoming flow conditions and the digital elevation model; combine the relative elevation, surface roughness and the cumulative characteristics of the incoming flow channel to construct the multi-channel terrain condition tensor.
[0025] Here, the relative elevation channel is obtained based on the local minimum elevation. Then for each grid location In the sampling set determined according to the direction of the incoming flow The positive elevation difference within the channel is statistically analyzed and accumulated according to distance attenuation weights. The resulting two-dimensional field is then uniformly normalized to obtain the cumulative characteristics of the incoming flow channel. Its representative expression is as follows:
[0026]
[0027] in, Indicates grid position The normalized cumulative characteristic value of the incoming flow channel; This represents the set of sampling grids upstream, downstream, or a combination of both at this location, determined by the direction of the incoming flow d; r represents any grid point in the sampling set; This represents the elevation at sampling point r. Indicates the current location elevation; This represents the relative distance between the sampling point r and the current position along the channel direction; H This represents the feature length used to control the distance decay scale; Indicates the distance decay weight; This represents a normalization operator that scales the accumulated results according to the uniform statistics used in the training phase. Accumulating only positive elevation differences highlights the continuous uplift effect along the direction of the incoming flow, while distance attenuation weights are used to reduce the excessive influence of distant terrain on the current location. The sampling set can be set as the upstream channel, the downstream channel, or a combination of both, and the weighting function can be replaced with linear, exponential, or piecewise attenuation forms. Finally, the relative elevation... Incoming flow channel cumulative characteristics and by The obtained normalized roughness channel Concatenate them into a multi-channel terrain condition tensor according to the channel dimension. .
[0028] In this application, by constructing relative elevation channels, inflow channels, cumulative feature channels, and roughness channels, it is possible to explicitly characterize topographic undulations, continuous uplift and shading relationships along the inflow direction, and the influence of surface roughness. Combined with multi-scale frequency modulation coding, it is possible to simultaneously capture long-range correlations across ridges and valleys, as well as local high-frequency features such as slope top acceleration, canyon jets, shear layers, and separation reattachment, thereby improving the prediction accuracy of wind fields in complex mountainous areas.
[0029] S102: Input the multi-channel terrain condition tensor into the pre-trained wind field evaluation model, perform multi-scale frequency modulation coding on the multi-channel terrain condition tensor based on the encoder to obtain shared coding features, and decode the shared coding features and the injected target height based on the decoder to obtain the multi-channel wind field prediction result.
[0030] In this step, the multi-channel terrain condition tensor is input into the wind field assessment model. The encoder performs multi-scale frequency modulation coding on the multi-channel terrain condition tensor to obtain shared coding features. The decoder decodes the shared coding features and the injected target height to obtain the multi-channel wind field prediction results.
[0031] The multi-channel wind field prediction results include pressure disturbance, turbulent kinetic energy, and wind speed in three directions.
[0032] Here, the encoder is a Transformer encoder with attention inversion and frequency domain dynamic scaling, used to extract global and local features on the M×M coding grid to obtain coding features; the decoder includes independent decoding branches set for five physical quantities, and injects the target height through feature linear modulation at the low and high decoding scales, and performs decoding based on the coding features and the target height to obtain a five-channel prediction result at the M×M coding grid scale.
[0033] In one possible implementation, the step of obtaining shared coding features by multi-scale frequency modulation coding of the multi-channel terrain condition tensor based on the encoder includes: A: The multi-channel terrain condition tensor is converted into M×M encoded features through convolutional feature extraction and scale adaptation, while retaining both low-resolution and high-resolution terrain features.
[0034] Here, the multi-channel terrain condition tensor First, convolutional feature extraction and scale adaptation are performed to obtain an M×M coding grid, and low-resolution terrain features FL and high-resolution terrain features FH are retained for subsequent decoding and fusion.
[0035] B: Map the encoded features into a token sequence according to a preset patch and add position information.
[0036] Here, the M×M encoded features are mapped to a token sequence according to a preset patch partitioning method, and location-related information is added to convert the two-dimensional mountain features into a sequence representation suitable for Transformer encoding. The patch size can be set according to computing resources and target precision.
[0037] C: Perform joint encoding on the token sequence, including attention main response, supplementary residuals, and frequency domain dynamic scaling, to obtain shared coding features.
[0038] Here, the token sequence is jointly encoded by attention main response, supplementary residuals, and frequency domain dynamic scaling to obtain shared coding features.
[0039] In one possible implementation, the joint encoding of the token sequence, which includes attention main response, supplementary residuals, and frequency domain dynamic scaling, to obtain shared coding features, includes: a: Perform a linear mapping on the token sequence to obtain query features, key features, and value features, and calculate the attention matrix based on the query features and key features.
[0040] Here, the attention matrix is determined using the following formula. A :
[0041] Where Q, K, and V represent query features, key features, and value features, respectively. This represents the attention matrix calculated from Q and K. Indicates dimension.
[0042] b: Generate a dynamic gating system determined by the current token sequence, and fuse the attention master response and the supplementary residual according to the dynamic gating system to obtain the attention fusion feature.
[0043] Here, the attention fusion features are determined using the following formula. :
[0044] in, Indicates the current token sequence X The generated dynamic gating.
[0045] c: Perform frequency domain transformation on the token sequence, and dynamically scale the frequency domain transformed features according to the content-related routing weights. Then, perform inverse frequency domain transformation to map back to the spatial domain to obtain frequency modulation features.
[0046] Here, the frequency modulation characteristics are determined using the following formula. :
[0047] in, as well as They represent the inverse frequency domain transform and the frequency domain transform, respectively. Indicates frequency domain modulation group index, Indicates the first q Group frequency domain modulation parameters, (X) is the first q The content-related routing weights of the group frequency domain modulation parameters are dynamically generated from the current input feature X and are used to adaptively select or combine different frequency domain modulation parameters based on the content of the terrain sample.
[0048] d: The token sequence, the attention fusion feature, and the frequency modulation feature are weighted and combined according to the fusion coefficient to obtain the output feature of the current coding block. Multiple coding blocks are concatenated to determine the shared coding feature.
[0049] Here, the token sequence, attention fusion features, and frequency modulation features are weighted and combined according to the fusion coefficient to obtain the output features of the current coding block. Multiple coding blocks are concatenated to determine the shared coding features.
[0050] In the wind field proxy task, the overall uplift, flow around, and large-scale velocity redistribution caused by terrain typically exhibit a relatively smooth, large-scale structure, while slope-top acceleration, canyon jets, shear layers, separation and reattachment, local backflow, and vertical velocity abrupt changes show stronger locality and higher spatial frequency. This encoding enables the network to simultaneously retain both types of information within a unified feature space and dynamically adjust the frequency band response according to different terrain features.
[0051] It should be noted that, using token features as input, a multi-scale frequency modulation Transformer coding structure incorporating attention master responses, supplementary residuals, and frequency domain dynamic scaling is employed to jointly encode long-range correlations across ridges and valleys, as well as local slope details, yielding coded features E. The number of coding layers, the number of attention heads, the channel grouping method, and the frequency domain modulation parameters can be configured according to the training data scale and computational resources.
[0052] In one possible implementation, the decoding of the shared coded features and the injected target height based on the decoder to obtain multi-channel wind field prediction results includes: The shared coding features are input into five independent decoding branches. In the low-resolution stage, low-resolution terrain features are fused and a first set of height modulation parameters generated from the target height are injected. In the high-resolution stage, high-resolution terrain features are fused and a second set of height modulation parameters generated from the target height are injected. Each branch generates a prediction field. The prediction fields are then restored to high resolution and aligned to the engineering grid to obtain multi-channel wind field prediction results.
[0053] Here, the inputs during the decoding process include shared coding features and low-resolution terrain features retained during the engineering mesh adaptation stage. and high-resolution terrain features The target height h is first scaled according to the training height range, then converted into a scaling vector and bias vector matching the current number of decoded feature channels by a height conditional mapping network, and injected into the low-resolution scale L and high-resolution scale H respectively. A representative expression is:
[0054]
[0055]
[0056] In the formula, and These represent the lower and upper limits of the target height range for the training samples, respectively. 's' represents the target height after scaling; 's' represents the decoding scale. L1 is a low-scale number. H 1 represents a high scale. Represents the height-conditional mapping network at the corresponding scale; and These represent the channel scaling vector and bias vector generated from the target height, respectively; This represents the decoded features that have already incorporated terrain information at scale s; ⊙ denotes element-wise multiplication; This represents the features modulated by the target height condition. Low-scale FiLM is used to adjust the background distribution of the overall flow field as a function of height, while high-scale FiLM is used to adjust the attenuation or enhancement of local terrain responses such as slopes, ridges, and canyons as a function of height. The modulated features then continue to the output layer of each physical quantity decoding branch.
[0057] Specifically, independent decoding branches are established for pressure disturbance, turbulent kinetic energy, and wind speed components in the x, y, and z directions. Each branch shares the same coding features. E However, they have independent decoding parameters. Each decoder head first fuses in the low-resolution stage. E With low-resolution terrain features The first set of height modulation parameters is generated based on the target height h; then the features are gradually restored to an M×M scale, and high-resolution terrain features are fused again in the high-resolution stage. A second set of height modulation parameters is then injected, and the base field branch and the local residual branch jointly generate the M×M prediction field corresponding to the physical quantities. This process allows the height condition to directly participate in decoding, rather than performing height interpolation after the network output.
[0058] It should be noted that high-resolution restoration uses subpixel rearrangement, transposed convolution, interpolated convolution, or other learnable upsampling methods to restore the prediction field to the engineering grid, while maintaining a one-to-one correspondence between the five physical quantities and the original engineering terrain grid space.
[0059] In one possible implementation, each independent decoding branch generates the corresponding prediction field using a base-residual form:
[0060] in, For the category of the predicted field, , This represents the prediction result for the c-th category at the current decoding scale. The base field components represent a relatively smooth background distribution. This is a correction component formed due to the influence of local topography. These correspond to pressure disturbance, turbulent kinetic energy, wind speed components in the x, y, and z directions, respectively.
[0061] In this application, during the five-branch decoding process, height modulation parameters generated from the target height are injected at both the low-resolution and high-resolution stages, allowing the target height to directly participate in the decoding process, rather than performing height interpolation after the network output. Low-scale modulation is used to adjust the background distribution of the overall flow field as a function of height, while high-scale modulation is used to adjust the attenuation or enhancement of the local terrain response as a function of height, thereby improving the continuity and accuracy of wind field prediction at different target heights.
[0062] S103: Perform inverse normalization on the multi-channel wind field prediction results to obtain a rapid assessment of wind resources in complex terrain at the target height.
[0063] In this step, the multi-channel wind field prediction results are inversely normalized to obtain a rapid assessment of wind resources in complex terrain at the target height.
[0064] In one possible implementation, the wind field assessment model is determined through the following steps: (1): Obtain training samples, which include multi-channel terrain condition samples, target height samples, and the real field of five-channel physical quantities obtained by computational fluid dynamics.
[0065] Here, the digital elevation model Z, surface roughness R, and inflow condition parameters I of the target area are obtained. For newly created samples, a computational fluid dynamics (CFD) domain is established according to a unified inflow coordinate direction. Inflow velocity, inflow direction, turbulent boundary, and ground boundary conditions are set. After convergence calculation, five-channel horizontal slices of pressure disturbance p, turbulent kinetic energy k, and velocity components u, v, and w are extracted at one or more target heights h to obtain the original training sample D_raw. File integrity, CFD convergence, and outlier checks are performed on the training sample D_raw. The terrain and flow field are unified to the same coordinate direction according to the inflow direction, and an N×N engineering mesh is formed through resampling or pruning. Normalized statistics of each physical quantity are calculated only based on the training set, and the same rules are applied to the training, validation, and online data accordingly.
[0066] (2): Perform channel-independent normalization on the real field of the five physical quantities to obtain the supervision data.
[0067] (3): Input the multi-channel terrain condition samples and target height samples into the neural network model to obtain the five-channel prediction results.
[0068] Here, the neural network model processes the multi-channel terrain condition samples and target height samples in the same way as the wind field assessment model described above, and will not be elaborated further here.
[0069] (4): Calculate the error between the five-channel prediction result and the supervision data using the loss function; update the network parameters of the neural network model in a unified manner according to the error, so that the shared encoder and each independent decoding branch can be trained together to determine the wind field evaluation model.
[0070] Here, the five-channel error between the five-channel prediction results and the supervised data is calculated using a loss function. Error backpropagation and parameter updates are then performed to enable synchronous learning of the shared encoder, frequency modulation unit, and five independent decoders. After each training round, forward prediction is performed on the validation set according to the same data processing rules as the training phase, and weights with better performance are selected based on preset evaluation metrics. This step outputs the weights of the trained candidate models and the validation results to determine the wind field assessment model.
[0071] The loss function employs one or more combinations of absolute error, squared error, and smoothed absolute error. Evaluation metrics include one or more of the following: full-field MAE, RMSE, absolute error of five channels, gradient error, structural similarity, ground truth-predicted field map comparison, and candidate camera position region error.
[0072] It should be noted that the selected model weights, input channel order, output variable order, normalized statistics, training height range, engineering mesh size, spatial resolution, terrain undulation statistical range, roughness statistical range, and incoming flow condition identifier I are saved to form a one-to-one corresponding deployment model package. Based on the training applicability range recorded in the deployment model package, the height range, terrain undulation, roughness, and incoming flow condition are checked. If within the range, a five-channel prediction field and / or derived results are output; if outside the range, an out-of-distribution identifier is given. For new ranges that need to be covered, CFD supervised samples are added, and the training step is repeated to update the deployment model package.
[0073] In this application, by combining offline CFD supervised sample training with online agent network forward inference, the online stage does not require providing the initial CFD flow field, external background wind field, or performing CFD iterative solutions, thus significantly improving the efficiency of wind resource assessment in complex mountainous areas. By constructing three-channel terrain conditions—relative elevation, inflow channel cumulative features, and roughness—and combining them with multi-scale frequency modulation coding, it can simultaneously capture large-scale flow correlations across ridges and valleys, as well as local high-frequency features such as slope top acceleration, canyon jets, and separation reattachment, improving the ability to characterize wind fields in complex terrains. By injecting target height conditions in the low-resolution and high-resolution decoding stages, the target height directly participates in decoding rather than being interpolated after output, improving the continuity and accuracy of predictions at different target heights. By sharing coding features and using five independent decoding heads, it achieves joint prediction and consistency constraints of pressure disturbances, turbulent kinetic energy, and wind speed components in three directions.
[0074] This application provides a method for rapid assessment of wind resources in complex mountainous areas. The method includes: constructing a multi-channel topographic condition tensor under preset incoming flow conditions; wherein the multi-channel topographic condition tensor includes at least relative elevation, surface roughness, and cumulative positive elevation difference features of the incoming flow channels; inputting the multi-channel topographic condition tensor into a pre-trained wind field assessment model; performing multi-scale frequency modulation encoding on the multi-channel topographic condition tensor based on an encoder to obtain shared encoded features; decoding the shared encoded features and the injected target height based on a decoder to obtain multi-channel wind field prediction results; wherein the multi-channel wind field prediction results include pressure disturbance, turbulent kinetic energy, and wind speeds in three directions; and performing inverse normalization on the multi-channel wind field prediction results to obtain a rapid assessment of wind resources in complex terrain at the target height. The online phase does not require providing the initial CFD flow field, external background wind field, or performing CFD iterative solutions, thus significantly improving the efficiency of wind resource assessment in complex mountainous areas. Through the collaborative representation of multi-channel terrain conditions and target height, combined with the encoder and decoder architecture, the assessment results of mountainous wind resources can be accurately determined with only terrain data and target height as input.
[0075] Please see Figure 2 , Figure 3 , Figure 2 This is one of the structural schematic diagrams of a rapid assessment device for complex mountain wind resources provided in the embodiments of this application; Figure 3 This is a second schematic diagram of a rapid assessment device for complex mountain wind resources provided in an embodiment of this application. Figure 2 As shown, the rapid assessment device 200 for complex mountain wind resources includes: The condition construction module 210 is used to construct a multi-channel terrain condition tensor under a preset incoming flow condition; wherein, the multi-channel terrain condition tensor includes at least relative elevation, surface roughness and cumulative positive elevation difference features of the incoming flow channel; The wind field assessment module 220 is used to input the multi-channel terrain condition tensor into a pre-trained wind field assessment model, perform multi-scale frequency modulation coding on the multi-channel terrain condition tensor based on the encoder to obtain shared coding features, and decode the shared coding features and the injected target height based on the decoder to obtain multi-channel wind field prediction results; wherein, the multi-channel wind field prediction results include pressure disturbance, turbulent kinetic energy and wind speed in three directions; The output module 230 is used to perform inverse normalization processing on the multi-channel wind field prediction results to obtain a rapid assessment of wind resources in complex terrain at the target height.
[0076] Furthermore, the condition construction module 210 constructs a multi-channel terrain condition tensor under the preset incoming flow conditions through the following steps: Obtain the digital elevation model and surface roughness data of the target area, and calculate the relative elevation based on the digital elevation model; Based on the preset incoming flow conditions and the digital elevation model, the cumulative characteristics of the incoming flow channel along the incoming flow direction are calculated; The relative elevation, surface roughness, and cumulative features of the incoming flow channels are combined to construct the multi-channel terrain condition tensor.
[0077] Furthermore, the wind field assessment module 220 is used to obtain shared coding features by performing multi-scale frequency modulation coding on the multi-channel terrain condition tensor based on the encoder: The multi-channel terrain condition tensor is converted into M×M encoded features through convolutional feature extraction and scale adaptation, while retaining both low-resolution and high-resolution terrain features. The encoded features are divided and mapped into a token sequence according to a preset patch, and position information is added; The token sequence is jointly encoded by attention main response, supplementary residuals, and frequency domain dynamic scaling to obtain shared coding features.
[0078] Furthermore, the wind field assessment module 220 is used to perform joint encoding on the token sequence, including attention main response, supplementary residuals, and frequency domain dynamic scaling, to obtain shared encoding features: A linear mapping is performed on the token sequence to obtain query features, key features, and value features. An attention matrix is then calculated based on the query features and key features. A dynamic gating system determined by the current token sequence is generated, and the attention master response and supplementary residual are fused according to the dynamic gating system to obtain attention fusion features; The token sequence is transformed in the frequency domain, and the features after the frequency domain transformation are dynamically scaled and grouped according to the content-related routing weights. Then, the frequency modulation features are mapped back to the spatial domain through inverse frequency domain transformation. The token sequence, the attention fusion feature, and the frequency modulation feature are weighted and combined according to the fusion coefficient to obtain the output feature of the current coding block. Multiple coding blocks are concatenated to determine the shared coding feature.
[0079] Furthermore, the wind field assessment module 220 is used to decode the shared encoded features and the injected target height based on the decoder to obtain multi-channel wind field prediction results: The shared coding features are input into five independent decoding branches respectively. In the low-resolution stage, low-resolution terrain features are fused and a first set of height modulation parameters generated by the target height are injected. In the high-resolution stage, high-resolution terrain features are fused and a second set of height modulation parameters generated by the target height are injected. Each branch generates a prediction field. Each predicted field is restored at high resolution and aligned to the engineering grid to obtain multi-channel wind field prediction results.
[0080] Furthermore, the wind field assessment module 220 is used to generate the corresponding prediction field for each independent decoding branch in the form of base field-residual:
[0081] in, For the category of the predicted field, , This represents the prediction result for the c-th category at the current decoding scale. The base field components represent a relatively smooth background distribution. This is a correction component formed due to the influence of local topography. These correspond to pressure disturbance, turbulent kinetic energy, wind speed components in the x, y, and z directions, respectively.
[0082] Furthermore, such as Figure 3 As shown, the rapid assessment device for wind resources in complex mountainous areas 200 also includes a training module 240, which determines the wind field assessment model through the following steps: Acquire training samples, which include multi-channel terrain condition samples, target height samples, and a five-channel physical quantity real field obtained through computational fluid dynamics; The real field of the five physical quantities is subjected to channel-independent normalization to obtain supervisory data; The multi-channel terrain condition samples and target height samples are input into the neural network model to obtain five-channel prediction results; The error between the five-channel prediction results and the supervision data is calculated using a loss function. The network parameters of the neural network model are updated uniformly based on the error, so that the shared encoder and each independent decoding branch are trained together to determine the wind field evaluation model.
[0083] This application provides a rapid assessment device for wind resources in complex mountainous areas. The device includes: a condition construction module for constructing a multi-channel topographic condition tensor under preset incoming flow conditions; wherein the multi-channel topographic condition tensor includes at least relative elevation, surface roughness, and cumulative positive elevation difference features of the incoming flow channels; a wind field assessment module for inputting the multi-channel topographic condition tensor into a pre-trained wind field assessment model, performing multi-scale frequency modulation encoding on the multi-channel topographic condition tensor based on an encoder to obtain shared encoding features, and decoding the shared encoding features and the injected target height based on a decoder to obtain multi-channel wind field prediction results; wherein the multi-channel wind field prediction results include pressure disturbance, turbulent kinetic energy, and wind speeds in three directions; and an output module for performing inverse normalization processing on the multi-channel wind field prediction results to obtain a rapid assessment of wind resources in complex terrain at the target height. The online phase does not require providing the initial CFD flow field, external background wind field, or performing CFD iterative solutions, thus significantly improving the efficiency of wind resource assessment in complex mountainous areas. Through the collaborative representation of multi-channel terrain conditions and target height, combined with the encoder and decoder architecture, the assessment results of mountainous wind resources can be accurately determined with only terrain data and target height as input.
[0084] Please see Figure 4 , Figure 4 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application. Figure 4 As shown, the electronic device 400 includes a processor 410, a memory 420, and a bus 430.
[0085] The memory 420 stores machine-readable instructions executable by the processor 410. When the electronic device 400 is running, the processor 410 communicates with the memory 420 via the bus 430. When the machine-readable instructions are executed by the processor 410, they can perform the operations described above. Figure 1 The steps of the rapid assessment method for complex mountain wind resources in the method embodiment shown are described in detail in the method embodiment, and will not be repeated here.
[0086] This application also provides a computer-readable storage medium storing a computer program, which, when executed by a processor, can perform the above-described actions. Figure 1 The steps of the rapid assessment method for complex mountain wind resources in the method embodiment shown are described in detail in the method embodiment, and will not be repeated here.
[0087] Those skilled in the art will understand that, for the sake of convenience and brevity, the specific working processes of the systems, devices, and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.
[0088] In the several embodiments provided in this application, it should be understood that the disclosed systems, apparatuses, and methods can be implemented in other ways. The apparatus embodiments described above are merely illustrative. For example, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. Furthermore, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Additionally, the shown or discussed mutual couplings, direct couplings, or communication connections may be through some communication interfaces; indirect couplings or communication connections between devices or units may be electrical, mechanical, or other forms.
[0089] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.
[0090] In addition, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit.
[0091] If the aforementioned functions are implemented as software functional units and sold or used as independent products, they can be stored in a processor-executable, non-volatile, computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or a portion of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.
[0092] Finally, it should be noted that the above-described embodiments are merely specific implementations of this application, used to illustrate the technical solutions of this application, and not to limit them. The scope of protection of this application is not limited thereto. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that any person skilled in the art can still modify or easily conceive of changes to the technical solutions described in the foregoing embodiments, or make equivalent substitutions for some of the technical features, within the scope of the technology disclosed in this application. Such modifications, changes, or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application, and should all be covered within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.
Claims
1. A rapid assessment method for wind resources in complex mountainous areas, characterized in that, The rapid assessment method for wind resources in complex mountainous areas includes: Construct a multi-channel terrain condition tensor under a preset incoming flow condition; wherein, the multi-channel terrain condition tensor includes at least relative elevation, surface roughness, and cumulative positive elevation difference characteristics of the incoming flow channel; The multi-channel terrain condition tensor is input into a pre-trained wind field evaluation model. The multi-scale frequency modulation encoding of the multi-channel terrain condition tensor is performed on the encoder to obtain shared encoding features. The shared encoding features and the injected target height are decoded on the decoder to obtain the multi-channel wind field prediction results. The multi-channel wind field prediction results include pressure disturbance, turbulent kinetic energy and wind speed in three directions. The multi-channel wind field prediction results are inversely normalized to obtain a rapid assessment of wind resources in complex terrain at the target height. The process of obtaining shared coding features by multi-scale frequency modulation coding of the multi-channel terrain condition tensor based on the encoder includes: The multi-channel terrain condition tensor is converted into M×M encoded features through convolutional feature extraction and scale adaptation, while retaining both low-resolution and high-resolution terrain features. The encoded features are divided and mapped into a token sequence according to a preset patch, and position information is added; The token sequence is jointly encoded by attention main response, supplementary residuals, and frequency domain dynamic scaling to obtain shared coding features.
2. The rapid assessment method for wind resources in complex mountainous areas according to claim 1, characterized in that, The following steps are used to construct a multi-channel terrain condition tensor under a preset inflow condition: Obtain the digital elevation model and surface roughness data of the target area, and calculate the relative elevation based on the digital elevation model; Based on the preset incoming flow conditions and the digital elevation model, the cumulative characteristics of the incoming flow channel along the incoming flow direction are calculated; The relative elevation, surface roughness, and cumulative features of the incoming flow channels are combined to construct the multi-channel terrain condition tensor.
3. The rapid assessment method for wind resources in complex mountainous areas according to claim 1, characterized in that, The joint encoding of the token sequence, including attention main response, supplementary residuals, and frequency domain dynamic scaling, yields shared encoding features, including: A linear mapping is performed on the token sequence to obtain query features, key features, and value features. An attention matrix is then calculated based on the query features and key features. A dynamic gating system determined by the current token sequence is generated, and the attention master response and supplementary residual are fused according to the dynamic gating system to obtain attention fusion features; The token sequence is transformed in the frequency domain, and the features after the frequency domain transformation are dynamically scaled and grouped according to the content-related routing weights. Then, the frequency modulation features are mapped back to the spatial domain through inverse frequency domain transformation. The token sequence, the attention fusion feature, and the frequency modulation feature are weighted and combined according to the fusion coefficient to obtain the output feature of the current coding block. Multiple coding blocks are concatenated to determine the shared coding feature.
4. The rapid assessment method for wind resources in complex mountainous areas according to claim 1, characterized in that, The process of decoding the shared encoded features and the injected target height based on the decoder to obtain multi-channel wind field prediction results includes: The shared coding features are input into five independent decoding branches respectively. In the low-resolution stage, low-resolution terrain features are fused and a first set of height modulation parameters generated by the target height are injected. In the high-resolution stage, high-resolution terrain features are fused and a second set of height modulation parameters generated by the target height are injected. Each branch generates a prediction field. Each predicted field is restored at high resolution and aligned to the engineering grid to obtain multi-channel wind field prediction results.
5. The rapid assessment method for wind resources in complex mountainous areas according to claim 4, characterized in that, Each independent decoding branch generates the corresponding prediction field using a base field-residual form: in, For the category of the predicted field, , This represents the prediction result for the c-th category at the current decoding scale. The base field components represent a relatively smooth background distribution. This is a correction component formed due to the influence of local topography. These correspond to pressure disturbance, turbulent kinetic energy, wind speed components in the x, y, and z directions, respectively.
6. The rapid assessment method for wind resources in complex mountainous areas according to claim 1, characterized in that, The wind field assessment model is determined through the following steps: Acquire training samples, which include multi-channel terrain condition samples, target height samples, and a five-channel physical quantity real field obtained through computational fluid dynamics; The real field of the five physical quantities is subjected to channel-independent normalization to obtain supervisory data; The multi-channel terrain condition samples and target height samples are input into the neural network model to obtain five-channel prediction results; The error between the five-channel prediction results and the supervision data is calculated using a loss function. The network parameters of the neural network model are updated uniformly based on the error, so that the shared encoder and each independent decoding branch are trained together to determine the wind field evaluation model.
7. A rapid assessment device for wind resources in complex mountainous areas, characterized in that, The rapid assessment device for wind resources in complex mountainous areas includes: The condition construction module is used to construct a multi-channel terrain condition tensor under a preset incoming flow condition; wherein, the multi-channel terrain condition tensor includes at least relative elevation, surface roughness, and cumulative positive elevation difference features of the incoming flow channel; The wind field assessment module is used to input the multi-channel terrain condition tensor into a pre-trained wind field assessment model, perform multi-scale frequency modulation encoding on the multi-channel terrain condition tensor based on the encoder to obtain shared encoding features, and decode the shared encoding features and the injected target height based on the decoder to obtain multi-channel wind field prediction results; wherein, the multi-channel wind field prediction results include pressure disturbance, turbulent kinetic energy and wind speed in three directions; The output module is used to perform inverse normalization processing on the multi-channel wind field prediction results to obtain a rapid assessment of wind resources in complex terrain at the target height. When the wind field assessment module obtains shared coding features by performing multi-scale frequency modulation coding on the multi-channel terrain condition tensor based on the encoder, the wind field assessment module is also used for: The multi-channel terrain condition tensor is converted into M×M encoded features through convolutional feature extraction and scale adaptation, while retaining both low-resolution and high-resolution terrain features. The encoded features are divided and mapped into a token sequence according to a preset patch, and position information is added; The token sequence is jointly encoded by attention main response, supplementary residuals, and frequency domain dynamic scaling to obtain shared coding features.
8. An electronic device, characterized in that, include: The device includes a processor, a memory, and a bus. The memory stores machine-readable instructions executable by the processor. When the electronic device is running, the processor communicates with the memory via the bus. The machine-readable instructions are executed by the processor to perform the steps of the rapid assessment method for complex mountain wind resources as described in any one of claims 1 to 6.
9. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed by a processor, performs the steps of the rapid assessment method for complex mountain wind resources as described in any one of claims 1 to 6.
