A smart wind field forecasting system and method based on multi-source data fusion

Abstract

This invention relates to the field of meteorological forecasting and discloses an intelligent wind field forecasting system and method based on multi-source data fusion. The method includes: granular processing of multi-source data, mapping the data to a micro-region multi-dimensional grid space; cross-projecting historical short-term wind field disturbance patterns and real-time observation features onto a dynamic hierarchical disturbance feature set, using multi-modal feature encoding and disturbance weight modulation; scanning transient disturbance features and constructing a hierarchical interaction strategy based on disturbance intensity and neighborhood information to adaptively adjust the contribution order and weighting coefficients of different data sources during the fusion process; performing multi-round dynamic feedback on the semi-serialized disturbance fusion representation in the micro-region grid network, with each iteration fine-tuning the grid weights based on local disturbance trends, neighborhood influences, and disturbance factor optimization results; and calculating the cumulative effect of local fluctuations in the fused feature matrix after dynamic feedback convergence. This invention improves the accuracy of predicting short-term fluctuations in micro-region wind fields.

Description

Technical Field

[0001] This invention relates to the field of meteorological forecasting, specifically to an intelligent wind field forecasting system and method based on multi-source data fusion. Background Technology

[0002] There are still significant shortcomings in predicting small-scale wind field changes in urban microclimate regions, especially short-term gusts around high-rise building clusters. Traditional meteorological stations are spaced far apart, making it difficult to obtain real-time wind speed and direction changes in local areas. Conventional numerical weather prediction models suffer from high computational costs and slow update cycles in high-resolution short-term predictions, making it difficult to capture micro-disturbances in the wind field in a timely manner. In specific application scenarios, such as low-altitude drone flights in large urban building clusters or safety monitoring of high-altitude wind turbine blades, sudden local gusts may cause flight path deviations or excessive mechanical stress, and existing systems cannot provide sufficiently detailed early warning information. Therefore, it is necessary to design an intelligent wind field forecasting system and method based on multi-source data fusion to improve the accuracy of predicting short-term fluctuations in micro-regional wind fields. Summary of the Invention

[0003] To address the shortcomings of existing technologies, this invention provides a wind field intelligent forecasting system and method based on multi-source data fusion, which has the advantage of improving the prediction accuracy of short-term fluctuations in micro-region wind fields and solves the problems mentioned in the background technology.

[0004] To achieve the aforementioned goal of improving the accuracy of predicting short-term fluctuations in micro-region wind fields, this invention provides the following technical solution: a smart wind field forecasting method based on multi-source data fusion, comprising the following steps: The multi-source data is granularized and mapped to a micro-region multi-dimensional grid space. A dynamic hierarchical perturbation feature set is generated according to local flow sensitivity and time continuity. Historical short-term wind field disturbance patterns and real-time observation features are cross-projected onto a dynamic hierarchical disturbance feature set. Through multimodal feature encoding and disturbance weight modulation, transient disturbance features with adaptive response capabilities are formed. Transient perturbation features are scanned to identify high-impact anomalies. A hierarchical interaction strategy is constructed based on perturbation intensity and neighborhood information to adaptively adjust the contribution order and weighting coefficients of different data sources in the fusion process, forming a semi-serialized perturbation fusion representation. The semi-serialized perturbation fusion representation is carried out in a micro-region grid network for multiple rounds of dynamic feedback. In each iteration, the grid weights are fine-tuned according to the local perturbation trend, neighborhood influence and perturbation factor optimization results to form a fusion feature matrix, and weighted correction is performed on potential anomalies and predicted unstable areas. After the dynamic feedback converges, the cumulative effect of local fluctuations in the fused feature matrix is ​​calculated to generate a global short-term wind field disturbance distribution map, and the wind speed, wind direction and disturbance probability prediction results for the next few minutes are output.

[0005] Preferably, the process of generating a dynamic hierarchical perturbation feature set based on local flow sensitivity and temporal continuity is as follows: By performing granular sampling of radar, satellite, ground observation station and numerical model data, a spatiotemporal index is established for various types of observation data in a micro-area multi-dimensional grid space; Based on local wind speed gradient, wind direction change rate and historical disturbance patterns, the flow sensitivity coefficient of each micro-grid point is calculated. Data is hierarchically archived according to temporal continuity and local sensitivity, generating a dynamic hierarchical perturbation feature set for cross-projection.

[0006] Preferably, the process of cross-projecting historical short-term wind field disturbance patterns and real-time observation features into a dynamic hierarchical disturbance feature set is as follows: The generated dynamic hierarchical perturbation feature set is used as input, and the micro-area perturbation features are aligned with historical short-term perturbation patterns according to the time window. By using feature similarity measurement methods, the mapping relationship between historical disturbance patterns and real-time observation features is identified within a dynamic hierarchical disturbance feature set, forming micro-area disturbance matching pairs; A cross-projection matrix is ​​constructed by stitching together the historical and real-time perturbation features of each micro-region grid point in the dynamic hierarchical perturbation feature set with multi-dimensional vectors.

[0007] Preferably, the process of forming transient disturbance characteristics with adaptive response capability is as follows: A multimodal feature coding network is used to extract nonlinear features from the cross-projection matrix, and historical perturbation patterns, real-time observation features and micro-area sensitivity coefficients are embedded using tensor quantization. Based on the perturbation weight modulation mechanism, the contribution ratio of different data sources in the micro-area feature representation is adaptively adjusted, and the micro-area transient perturbation features with adaptive response capability are output.

[0008] Preferably, the process of constructing a hierarchical interaction strategy based on perturbation strength and neighborhood information is as follows: Threshold scanning and local gradient analysis are performed on the transient disturbance characteristics of micro-regions to identify abnormal disturbance amplitudes and high-impact areas. By combining the perturbation intensity, temporal continuity, and sensitivity coefficient of outliers with those of neighboring grid points, a hierarchical interaction strategy is constructed to guide the contribution order and weight modulation of each data source.

[0009] Preferably, the process of forming a semi-serialized perturbation fusion representation is as follows: Based on a hierarchical interaction strategy, the perturbation intensity, neighborhood influence, and micro-region sensitivity of different observation sources are comprehensively evaluated, and the contribution order of each observation source in the fusion process is dynamically arranged according to the evaluation results. Initial weighting coefficients are assigned to each observation source according to the order of contribution, and the weighting coefficients are adaptively updated according to real-time perturbation changes. Highly sensitive micro-regions are given priority to enter the fusion sequence. Short-time response weighting is applied to high-weight micro-region features in the preceding contribution position, while long-term smoothing weighting is applied to low-weight micro-region features in the following contribution position, generating a semi-serialized perturbation fusion representation.

[0010] Preferably, the process of fusing the semi-serialized perturbation representation in a micro-grid network and performing multiple rounds of dynamic feedback is as follows: The semi-serialized perturbation fusion representation is input into the micro-grid network, and information is transferred and local features are updated according to the neighborhood influence relationship. In each round of feedback, the grid weights are adaptively fine-tuned by combining the micro-region perturbation trend and transient characteristics to form an optimized fusion representation.

[0011] Preferably, the process of forming the fused feature matrix is ​​as follows: The weights of micro-region grid points are adjusted based on the micro-region disturbance change trend and the differences in the characteristics of neighboring grid points in each round of dynamic feedback. By combining the optimization results of the disturbance factor, the weights of high-risk or unstable areas are enhanced or suppressed to form a fusion weight matrix, which fully reflects the cumulative effect of short-term wind field disturbance in micro-areas.

[0012] Preferably, the process of generating a global short-term wind field disturbance distribution map is as follows: The cumulative effect of local perturbations at each micro-region grid point in the fused feature matrix is ​​calculated. By combining grid point weights, historical patterns, and neighborhood sensitivity, a comprehensive assessment and weighted correction is made for short-term wind speed, wind direction, and disturbance probability. Output the generated global short-term wind field disturbance distribution map and mark potentially high-risk areas on the map.

[0013] A wind field intelligent forecasting system based on multi-source data fusion includes: Particle processing module: Performs granular processing on multi-source wind field data and maps it to a micro-area multi-dimensional grid space to generate a dynamic hierarchical perturbation feature set; The perturbation coding module cross-projects historical short-term wind field perturbation patterns with real-time observation features, and forms adaptive transient perturbation features through multimodal feature coding and weight modulation. Anomaly scheduling module: Scans transient disturbance features, identifies high-impact anomalies, adaptively adjusts the fusion weights of different data sources based on disturbance intensity and neighborhood information, and generates a semi-serialized disturbance fusion representation; Feedback optimization module: Performs multiple rounds of dynamic feedback on the semi-serialized fusion representation in the micro-grid network, fine-tunes the grid weights and corrects potential anomalies, and forms a fusion feature matrix; Global Prediction Module: Calculates the cumulative effect of local disturbances in the fused feature matrix, generates the global short-term wind field disturbance distribution, and outputs the wind speed, wind direction, and disturbance probability prediction for the next few minutes.

[0014] Compared with existing technologies, this invention provides a wind field intelligent forecasting system and method based on multi-source data fusion, which has the following beneficial effects: This invention achieves a refined characterization of wind field disturbance features by granularizing multi-source heterogeneous data and combining it with micro-region multidimensional grid spatial mapping and local flow sensitivity and temporal continuity analysis, enabling high-resolution capture and characterization of short-term wind field changes. Utilizing cross-projection of historical short-term wind field disturbance patterns with real-time observation features, as well as multimodal feature encoding and adaptive modulation of disturbance weights, the invention endows micro-region transient disturbance features with a high degree of adaptive response capability, thereby achieving effective information integration and redundancy reduction among data sources. Furthermore, through a hierarchical interaction strategy based on disturbance intensity and neighborhood information, the contribution order and weighting of different data sources are dynamically adjusted. This method employs a semi-serialized perturbation fusion representation to improve prediction accuracy and anomaly detection capabilities. Multiple rounds of dynamic feedback iteration are implemented within a micro-grid network, incorporating local perturbation trends, neighborhood influences, and perturbation factors into the grid weights for fine-tuning, forming a fused perturbation feature matrix. This matrix effectively weights and corrects potential anomalies and unstable prediction areas, enhancing the robustness and reliability of short-term wind field forecasts. A global short-term wind field perturbation distribution map is generated through cumulative effect calculations, enabling accurate predictions of wind speed, direction, and perturbation probability. High-risk areas are also marked, providing quantifiable and visualized decision support information for wind power dispatching, aviation navigation, disaster prevention and mitigation, and urban wind environment management. This method demonstrates significant advantages in refined feature capture, real-time adaptability, anomaly detection, prediction accuracy, and risk visualization, significantly improving the intelligence and practical value of short-term wind field forecasting. Attached Figure Description

[0015] Figure 1 This is a schematic diagram of the method of the present invention; Figure 2 This is a schematic diagram of the structure of the present invention. Detailed Implementation

[0016] 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.

[0017] Example 1: Please refer to Figure 1 As shown in the figure, a wind field intelligent forecasting method based on multi-source data fusion in an embodiment of the present invention includes the following steps: S1: Perform granular processing on multi-source data, map the data to a micro-region multi-dimensional grid space, and generate a dynamic hierarchical perturbation feature set according to local flow sensitivity and time continuity.

[0018] The process of generating a dynamic hierarchical perturbation feature set in S1 based on local flow sensitivity and temporal continuity is as follows: By granular sampling of radar, satellite, ground observation station, and numerical model data, a spatiotemporal index is established for various observation data within a micro-area multidimensional grid space. Three-dimensional echo volumes from phased array radar, high-frequency wind speed and direction records from ground observation stations, and short-term forecast fields provided by numerical models are granularly sampled with a unified time step and spatial resolution. During granular sampling, by setting grid size, time step width, and noise suppression threshold, the original continuous field data is discretized into processable micro-area data blocks. Each micro-area block is assigned an index label including timestamp, spatial coordinates, observation source, variable type, and sampling density. Based on these index labels, a micro-area multidimensional grid space is constructed, enabling data from different sources to be comparable and searchable within this space. Based on local wind speed gradient, wind direction change rate, and historical disturbance patterns, the flow sensitivity coefficient of each micro-region grid point is calculated. On the basis of a granular structure, key variables such as local wind speed gradient, wind direction change rate, and turbulent kinetic energy are extracted for each micro-region grid point. Difference analysis and short-term stability assessment are performed on the time series of these variables to describe the strength of the micro-region's wind field response to external disturbances. Short-term wind field disturbance patterns corresponding to the location of the micro-region are retrieved from the historical database. Using disturbance intensity, duration, and disturbance propagation rate from the model library, the local environmental sensitivity at the current moment is corrected. After the real-time sensitive variables and historical disturbance reference quantities are fused in a weighted manner, the flow sensitivity coefficient of each micro-region grid point is calculated to characterize the importance ranking and response priority of different micro-regions in the subsequent fusion stage. Data is archived in layers according to temporal continuity and local sensitivity to generate a dynamic layered perturbation feature set for cross-projection. After obtaining the flow sensitivity coefficient, the feature sequences of the same micro-region in a continuous time window are subjected to stability analysis to identify their temporal continuity structure. The data is divided into a short-term rapid change layer, a medium-term transition layer, and a stable background layer according to the continuity level. Referring to the flow sensitivity coefficient of each micro-region, data segments with high sensitivity are preferentially placed in the upper feature layer, and data segments with low sensitivity are placed in the lower layer, thus forming a layered system that interweaves temporal continuity layering and sensitivity level layering. All data from all sources are archived according to the above two-dimensional standard to construct a multi-level, indexable, and spatiotemporally comparable dynamic layered perturbation feature set.

[0019] S2: Cross-project historical short-term wind field disturbance patterns and real-time observation features into a dynamic hierarchical disturbance feature set, and form transient disturbance features with adaptive response capabilities through multimodal feature encoding and disturbance weight modulation.

[0020] The process of cross-projecting historical short-term wind field disturbance patterns and real-time observation features into a dynamic hierarchical disturbance feature set in S2 is as follows: The generated dynamic hierarchical perturbation feature set is used as input, and the micro-area perturbation features are aligned with historical short-term perturbation patterns according to a time window. The obtained dynamic hierarchical perturbation feature set is used as the basic input data for cross-projection. By reading the perturbation sequences and their corresponding time labels of different micro-area grid points in the feature set, these real-time observed perturbation features are aligned with the pre-built historical short-term perturbation pattern library by time window. The alignment process includes: setting a sliding time window, establishing time anchors for real-time sequences, matching the time periods of historical patterns, and normalizing the time scale of both to eliminate differences in dimensions such as sampling frequency and data timeliness. To ensure alignment accuracy, boundary correction and missing data imputation are also performed on the data on both sides of the window, so that the real-time perturbation features of each micro-area can establish a stable and comparable correspondence with historical patterns within the same time interval. This paper utilizes feature similarity measurement methods to identify the mapping relationship between historical disturbance patterns and real-time observed features within a dynamic hierarchical disturbance feature set, forming micro-area disturbance matching pairs. The time-aligned data is input into the feature similarity analysis module, which employs multi-index similarity measurement methods, including Euclidean distance, dynamic time warping, disturbance trend angle difference, and gradient matching index, to perform grid-by-grid comparison between historical disturbance patterns and real-time micro-area disturbance features in the dynamic hierarchical disturbance feature set. During the comparison, the system not only calculates the overall similarity of the feature sequences but also performs weighted analysis on the local matching degree of disturbance peak positions, disturbance growth stages, and decay stages to identify the true disturbance correlation. Based on the similarity results, corresponding historical and real-time disturbance matching pairs are generated for each micro-area grid point, and disturbance correlation weights are assigned according to the strength of the matching pairs, forming a micro-area disturbance mapping structure. This enables disturbance information from different sources and time periods to establish a traceable and projectable mapping framework within the same feature set. A cross-projection matrix is ​​constructed by concatenating the historical and real-time disturbance features of each micro-region grid point in the dynamic hierarchical disturbance feature set with multi-dimensional vectors. After identifying the disturbance matching pairs of each micro-region, the historical disturbance feature vectors and real-time disturbance feature vectors associated with that micro-region are expanded dimensionally and concatenated in a multi-dimensional manner to form a joint feature vector containing various structured features such as wind speed, wind direction, disturbance gradient, and disturbance persistence. During the concatenation process, dimensional alignment, scale normalization, and feature weight redistribution are used to ensure that data from different sources remain comparable and numerically stable in the same vector space. The joint feature vectors of all micro-regions are arranged in spatial coordinate order to construct a cross-projection matrix, which is used to describe the mapping and projection effect of historical disturbance patterns under real-time observation conditions, preserving the historical disturbance structure while embedding the detailed differences of real-time observation.

[0021] The process of forming transient perturbation characteristics with adaptive response capability in S2 is as follows: A multimodal feature encoding network is employed to extract nonlinear features from the cross-projection matrix, embedding historical perturbation patterns, real-time observation features, and micro-area sensitivity coefficients into a tensor. The constructed cross-projection matrix serves as the multimodal input source, containing historical short-term perturbation patterns, real-time observation features, and micro-area sensitivity coefficients extracted from a dynamic hierarchical perturbation feature set. This matrix is ​​input into the multimodal feature encoding network, which is composed of a combination of convolutional layers, graph neural structures, or temporal attention layers. It can perform nonlinear transformations and deep semantic decoupling for data from different sources. During the encoding process, convolutional embedding and feature compression are performed on the spatial, temporal, and perturbation sensitivity dimensions of the matrix, respectively, so that the multimodal data is converted into a unified high-dimensional tensor representation. At the same time, the differences in historical perturbation structure, current observation response, and micro-area sensitivity are preserved. During training, the encoder automatically learns cross-modal coupling relationships based on perturbation change patterns, thereby forming a fusion tensor with structural and temporal consistency. Based on the perturbation weight modulation mechanism, the contribution ratio of different data sources in the micro-area feature representation is adaptively adjusted, outputting micro-area transient perturbation features with adaptive response capabilities. Based on the perturbation weight modulation mechanism, a dynamic weighting operation is performed on the tensor after multimodal encoding, thereby realizing the adaptive adjustment of the contribution of different data sources in the micro-area feature representation. According to the historical perturbation credibility, real-time observation priority, and micro-area sensitivity coefficient reflected in the cross-projection matrix, three modulation factors are initialized, and corresponding weight vectors are generated through a learnable weight generation module. The weight vectors are mapped to each modal dimension of the fusion tensor, and proportional modulation is applied to the sub-feature channels of different sources, so that the data source with higher weight has stronger feature dominance in the high-sensitivity region, while the data source with lower weight participates in the fusion in a complementary manner. Since the weight modulation mechanism can be dynamically updated with real-time wind field changes during model operation, the final output micro-area transient perturbation features can adaptively reflect the current wind field perturbation intensity, change rate, and sensitive area response capability, enabling the end-to-end prediction system to quickly adapt to complex wind field scenarios and achieve highly timely local perturbation response modeling.

[0022] S3: Scan transient disturbance features, identify high-impact anomalies, construct a hierarchical interaction strategy based on disturbance intensity and neighborhood information, adaptively adjust the contribution order and weighting coefficients of different data sources in the fusion process, and form a semi-serialized disturbance fusion representation.

[0023] The process of constructing a hierarchical interaction strategy in S3 based on perturbation strength and neighborhood information is as follows: Threshold scanning and local gradient analysis are performed on the transient disturbance characteristics of micro-regions to identify anomalous disturbance amplitudes and high-impact areas. A grid-by-grid threshold scan is implemented for the transient disturbance characteristics of micro-regions. The threshold is dynamically set based on the historical statistical distribution of disturbances (e.g., quantiles, mean plus a certain number of standard deviations, or based on empirical risk thresholds) to distinguish between normal fluctuations and anomalous amplitudes. Local gradient analysis is performed on candidate grid points that pass the threshold screening in both time and space, calculating the temporal derivative (i.e., instantaneous rate of change) and spatial derivative (i.e., the difference gradient with neighboring grid points) at each candidate point. And higher-order differences (such as second derivatives to identify acceleration or decay trends), and the suddenness and propagation of disturbances are evaluated based on these derivative information. To enhance detection robustness, short-time energy density index and frequency domain energy spectrum mutation detection can be introduced in parallel. Noise false alarms are reduced by joint time-frequency domain discrimination. Based on the results of threshold scanning and gradient analysis, grid points that deviate significantly from the background distribution in amplitude, transient rate or spectral characteristics are marked as disturbance amplitude anomalies. Spatially adjacent and temporally related anomalies are merged into high-impact regions through clustering or connected component analysis. Anomalies are evaluated in combination with their neighborhood grid points based on perturbation intensity, temporal continuity, and sensitivity coefficients. A hierarchical interaction strategy is constructed to guide the contribution order and weight modulation of various data sources. A combined evaluation is performed on each anomaly and its neighborhood grid point set. Evaluation metrics include: neighborhood perturbation intensity (represented by the perturbation amplitude and average energy density within the neighborhood grid points), temporal continuity metrics (measured by perturbation duration, persistence index, and autocorrelation coefficient within a sliding window), micro-area sensitivity coefficient (calculated from the previous dynamic stratification process), and historical perturbation recurrence rate (based on the frequency of similar events retrieved from a historical pattern database). Based on these metrics, a multi-metric approach is adopted. The system employs either fractional or multidimensional ranking methods (such as weighted linear scoring, scaling followed by stepwise screening, or hierarchical discriminant analysis) to prioritize outliers and categorizes them into several levels (e.g., high priority, medium priority, low priority). The hierarchical interaction strategy uses this level division as its core rule: for high-priority outliers, the strategy prioritizes real-time high-resolution observation sources and places them at the front of the fusion sequence, while simultaneously increasing the initial weighting coefficients of the corresponding data sources; for medium-priority outliers, the strategy uses a weighted parallel approach combining historical patterns and real-time observations; for low-priority outliers, the real-time triggering frequency is reduced, and background field smoothing is applied. In addition, the hierarchical interaction strategy includes neighborhood propagation rules (i.e., when multiple grid points in the neighborhood simultaneously reach a higher level, the contribution order is expanded according to spatial connectivity) and threshold triggering rules (i.e., when an indicator exceeds a higher threshold, its level is immediately increased). This strategy outputs a specific contribution order table and initial weighting coefficient suggestions for the data sources, providing them in an updatable parameterized format to the subsequent weight modulation module for adaptive fusion order control and dynamic weighting adjustment.

[0024] The process of forming a semi-serialized perturbation fusion representation in S3 is as follows: A hierarchical interaction strategy is used to comprehensively evaluate the perturbation intensity, neighborhood influence, and micro-area sensitivity of different observation sources. The contribution order of each observation source in the fusion process is dynamically ranked according to the evaluation results. The constructed hierarchical interaction strategy is invoked, using the micro-area perturbation intensity, neighborhood influence, and sensitivity coefficient of each observation source (such as radar echo, satellite brightness temperature, ground wind speed and direction observations, and numerical model forecast fields) as core evaluation indicators. For the performance of each observation source in the same micro-area, the perturbation energy density, perturbation direction consistency, temporal continuity index, and propagation gradient of the observation source in the neighborhood grid points at the current moment are calculated to quantify the perturbation intensity, neighborhood influence, and sensitivity coefficient of each observation source. The ability of observation sources to contribute to local disturbances in real time, combined with the micro-area sensitivity coefficient in the dynamic stratification process, gives higher weight to micro-areas that are easily triggered or more sensitive to short-term wind field changes, so as to more accurately identify key disturbance sources. The above indicators are comprehensively scored by weighted linear combination or multi-indicator ranking method. Those with higher scores are given priority in the fusion sequence, which guides the formation of a dynamically arranged order of observation source contributions. This order of contributions is not fixed, but is updated in real time according to the disturbance changes within the continuous sampling window, ensuring that the fusion process can flexibly respond to the rapidly changing wind field structure, so that the importance of information contribution matches the actual disturbance situation at any time. Initial weighting coefficients are assigned to each observation source according to its contribution order, and the weighting coefficients are adaptively updated based on real-time perturbation changes. Highly sensitive micro-regions are given priority in the fusion sequence. After obtaining the dynamic contribution order, initial weighting coefficients are assigned to each observation source according to this order. A proportional allocation based on reverse order, Softmax weight mapping, or a parameterized assignment mode based on an adaptive factor for perturbation intensity is used to ensure that observation sources with higher contribution orders have higher initial weights. The weights of each observation source are updated in real-time adaptively, and the update mechanism can be based on the rate of change of perturbation intensity, neighborhood coupling trend, and micro-region... Sensitivity changes and short-term historical error backtracking are corrected through time-recursive update formulas or sliding window-based adjustment strategies. When a micro-region is at a high sensitivity level or encounters a sudden disturbance event, the weight of its corresponding observation source will be increased by accelerating the update, so that the micro-region and its related data will be given priority in entering the fusion sequence. Conversely, for observation sources with stable disturbances, low sensitivity, or weakened contribution in the short term, their weights will be gradually reduced according to a slow reduction strategy. This adaptive weighting mechanism ensures that the fusion model can optimize the importance ranking of data sources in real time as the external field changes, and realize dynamic response to rapidly changing wind field structures. High-weight micro-region features at earlier contribution positions are weighted using short-time response, while low-weight micro-region features at later contribution positions are weighted using long-term smoothing, generating a semi-serialized perturbation fusion representation. For micro-region features at the forefront of the contribution order and with weights greater than a set threshold, a short-time response weighting method is used, i.e., a time-sensitive weighting factor is introduced to allow these features to quickly reflect local perturbation changes in the fusion matrix. This weighting typically includes an exponential time decay kernel, a fast response filter, or a time derivative-based enhancement term. For micro-region features at later contribution positions... For the feature, a long-term smoothing weighting strategy is adopted to reduce the impact of short-term noise fluctuations. This strategy usually introduces moving average, low-frequency filtering, long-window smoothing kernel or steady-state feature preservation term to keep the background wind field structure continuous and stable during fusion. The high-weight features after short-term response weighting and the low-weight features after long-term smoothing weighting are fused in a serialized but not completely time-dependent manner. While retaining the ability to sensitively capture local instantaneous disturbances, the stability of the background field is also taken into account, forming a semi-serialized perturbation fusion representation that reflects the sorting logic and retains some parallel processing characteristics.

[0025] S4: The semi-serialized perturbation fusion representation is dynamically fed back in a micro-region grid network in multiple rounds. In each iteration, the grid weights are fine-tuned based on the local perturbation trend, neighborhood influence, and perturbation factor optimization results to form a fusion feature matrix, which is used to make weighted corrections for potential anomalies and predicted unstable regions.

[0026] In S4, the process of fusing semi-serialized perturbations into a multi-round dynamic feedback representation within a micro-grid network is as follows: The semi-serialized perturbation fusion representation is input into the micro-grid network, and information transmission and local feature updates are performed according to the neighborhood influence relationship. The obtained semi-serialized perturbation fusion representation is loaded into the micro-grid network one by one according to the micro-grid index. Each grid node saves its corresponding fusion feature vector, time stamp, initial weight, and perturbation uncertainty index. Based on the preset neighborhood structure (such as a fixed neighborhood radius filtered by geographical distance or a dynamic neighborhood determined adaptively according to grid density), an influence relationship graph between grids is constructed, enabling each grid to identify its own set of neighboring grids and the corresponding neighborhood influence degree. During the information transmission stage, each grid packs its current fusion feature, perturbation change trend, short-time response feature, and historical smoothing index into message units and sends them to neighboring grids according to the neighborhood relationship. After receiving the messages, the neighboring grids filter and weight the messages from different neighbors. Based on the confidence of the message source, distance decay coefficient, and neighborhood perturbation consistency, the information is differentiated, filtering out perturbation information with insufficient confidence or excessive directional deviation, and prioritizing stable features that are consistent with the physical evolution of the grid. After message aggregation is completed, the grid uses the aggregated multi-source perturbation information to update the local features of the current round, and generates new temporary feature representations containing the influence of neighborhood perturbations through nonlinear activation or feature reconstruction. In each round of feedback, the grid weights are adaptively fine-tuned by combining the micro-region perturbation trend and transient characteristics to form an optimized fusion representation. The micro-region grid network performs dynamic weight adjustment on each grid point based on the feature update results to enhance the network's sensitivity to local perturbation changes. At the beginning of each round of feedback, the system calculates the perturbation trend index of the grid point, including the direction, magnitude, and stability of the perturbation change in the most recent time slice, as well as whether it is consistent with the perturbation trend of neighboring grid points. At the same time, transient feature information, such as the response intensity of short-term abrupt changes, local gradient changes, and anomalous jump characteristics, is extracted. These indicators, along with the grid point's weights in the previous round, are used as input to adaptively fine-tune the current weights. The adjustment direction is mainly based on whether the perturbation is enhanced, whether the neighborhood consistency is improved, and whether the short-term perturbation is strengthened. Whether the response shows a sudden change with physical significance, the weights of grid points with rapidly increasing perturbations or high correlation with neighborhood evolution will be increased to strengthen their influence in subsequent iterations; the weights of grid points with unstable perturbations or low neighborhood consistency will be partially weakened to avoid amplifying the effect of local noise on the overall prediction. After the weight update is completed, the fusion feature of the grid point is recalculated according to the new weights, so that high-weight grid points contribute more short-term perturbation information in the fusion, while the weights of long-term smooth features of low-weight grid points are relatively increased, thereby realizing hierarchical dynamic correction of the fusion feature. Multiple rounds of feedback will continue to execute the above process. When the weight changes are below the threshold for several consecutive rounds or the uncertainty of the network perturbation tends to stabilize, the iteration process stops, forming the final optimized fusion representation.

[0027] The process of forming the fusion feature matrix in S4 is as follows: Based on the micro-region perturbation change trend and the differences in characteristics of neighboring grid points in each round of dynamic feedback, the weights of micro-region grid points are adjusted. At the end of each round of feedback, the perturbation change trend indicators of each micro-region grid point in this round (e.g., the direction of increase or decrease in amplitude, rate of change, and frequency of peak occurrence in the short term) and the characteristic differences with its neighboring grid points in the same time period (e.g., directional deviation, energy density difference, and inconsistency in spectral characteristics) are statistically recorded. Based on these quantitative results, a hierarchical decision-making process is adopted to adjust the grid point weights: for grid points that show a significant continuous increasing or decreasing trend and are highly consistent with the neighborhood, their weight share is increased in this round according to priority to amplify their contribution; for grid points that show isolated fluctuations or are significantly inconsistent with the neighborhood, their weights are reduced to suppress noise propagation. The weight adjustment is accompanied by several engineering safeguards, including: smoothing the weight changes to prevent abrupt changes (e.g., applying sliding window averaging or incremental limits), triggering backtracking checks for extreme abrupt changes and retaining historical snapshots to support rollback, and scaling the weight adjustment range according to the grid point uncertainty index to ensure that the weight changes of low-confidence grid points are more conservative. Based on the perturbation factor optimization results, the weights of high-risk or unstable areas are enhanced or suppressed to form a fused weight matrix that fully reflects the cumulative effect of short-term wind field perturbations in micro-regions. After completing the initial weight adjustment based on trends and neighborhood differences, the output of the perturbation factor optimization module is introduced as the basis for secondary correction. This module comprehensively considers the historical perturbation recurrence probability, current uncertainty assessment, external environmental interference indicators, and the model's inherent error distribution, providing risk intensity and stability suggestions for identified high-risk or unstable areas. Weight enhancement is implemented for areas judged as high-risk or highly unstable (to highlight their importance in the final fused representation). The system either enhances or suppresses the influence of the micro-region (to suppress possible false positives or noise amplification), while setting upper and lower limits for the magnitude of enhancement or suppression to avoid over-adjustment. After completing the secondary correction of all grid points, the final weights of all micro-region grid points are arranged according to spatial index and stored as a fusion weight matrix. The matrix not only records the weight values, but also includes the weight confidence, uncertainty measure and weight source label (e.g., trend adjustment, neighborhood correction, disturbance factor correction) for each grid point. This fusion weight matrix serves as a complete representation of the cumulative effect of short-term wind field disturbance in the micro-region and is used for weighted reconstruction of the fusion feature vector and global disturbance accumulation calculation.

[0028] S5: After the dynamic feedback converges, the cumulative effect of local fluctuations in the fused feature matrix is ​​calculated to generate a global short-term wind field disturbance distribution map, and the wind speed, wind direction and disturbance probability prediction results for the next few minutes are output.

[0029] The process of generating a global short-time wind field disturbance distribution map in S5 is as follows: The cumulative effect of local perturbations at each micro-region grid point in the fusion feature matrix is ​​calculated. The cumulative perturbation representation of the micro-region after multiple rounds of dynamic feedback, semi-serialized fusion and weight optimization is read, including the fusion feature value, micro-region weight and uncertainty index of each grid point. For each micro-region grid point, the cumulative effect is calculated based on its short-term perturbation response, long-term trend components and neighborhood influence. The grid point cumulative perturbation amount is formed by summarizing the perturbation amplitude, spectrum change and peak occurrence within a finite short-term window. At the same time, the abnormal jump points are denoised and corrected by combining neighborhood consistency to ensure that the cumulative result reflects the short-term perturbation characteristics of the micro-region and retains the continuity of spatial distribution. By combining grid point weights, historical patterns, and neighborhood sensitivity, short-term wind speed, wind direction, and disturbance probability are comprehensively evaluated and weighted for correction. After completing the micro-area cumulative disturbance representation, the cumulative disturbance value of each grid point is jointly analyzed with its micro-area weight, historical disturbance pattern matching degree, and neighborhood sensitivity index to achieve the mapping and correction of wind speed, wind direction, and disturbance probability. The cumulative disturbance is weighted and transformed according to the weights and sensitivity, and mapped to short-term wind speed and wind direction adjustment factors. The wind direction deviation of isolated abnormal grid points is corrected by using neighborhood consistency and historical patterns to maintain the continuity and physical rationality of the spatial field. The disturbance probability is calculated by weighting the cumulative disturbance, historical recurrence probability, and current uncertainty, and outputs the probability value of significant disturbance occurring at each grid point within the short-term window. The whole process ensures that the short-term wind field prediction reflects both the cumulative effect of micro-area disturbances and takes into account historical patterns and spatial physical consistency. Output the generated global short-term wind field disturbance distribution map and mark potential high-risk areas on the map; visualize the micro-area cumulative disturbance representation and the mapped wind speed, wind direction and disturbance probability information to generate a global short-term wind field disturbance distribution map. The map includes wind speed contour lines, wind direction vector field and disturbance probability thermal layer. The marking of high-risk areas is based on the disturbance probability threshold, cumulative disturbance intensity and historical high-risk event matching degree: when a micro-area or consecutive micro-area grid points simultaneously meet the high disturbance probability and historical high-risk pattern matching conditions, the area is marked as high-risk, and a risk level and confidence level description are attached.

[0030] Example 2: As Figure 2 As shown, a wind field intelligent forecasting system based on multi-source data fusion includes: Particle processing module: Performs granular processing on multi-source wind field data and maps it to a micro-area multi-dimensional grid space to generate a dynamic hierarchical perturbation feature set; The perturbation coding module cross-projects historical short-term wind field perturbation patterns with real-time observation features, and forms adaptive transient perturbation features through multimodal feature coding and weight modulation. Anomaly scheduling module: Scans transient disturbance features, identifies high-impact anomalies, and adaptively adjusts the fusion weights of different data sources based on disturbance intensity and neighborhood information to generate a semi-serialized disturbance fusion representation; Feedback optimization module: Performs multiple rounds of dynamic feedback on the semi-serialized fusion representation in the micro-grid network, fine-tunes the grid weights and corrects potential anomalies, and forms a fusion feature matrix; Global Prediction Module: Calculates the cumulative effect of local disturbances in the fused feature matrix, generates the global short-term wind field disturbance distribution, and outputs the wind speed, wind direction, and disturbance probability prediction for the next few minutes.

[0031] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0032] 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 smart wind field forecasting method based on multi-source data fusion, characterized in that, Includes the following steps: The multi-source data is granularized and mapped to a micro-region multi-dimensional grid space. A dynamic hierarchical perturbation feature set is generated according to local flow sensitivity and time continuity. Historical short-term wind field disturbance patterns and real-time observation features are cross-projected onto a dynamic hierarchical disturbance feature set. Through multimodal feature encoding and disturbance weight modulation, transient disturbance features with adaptive response capabilities are formed. Transient perturbation features are scanned to identify high-impact anomalies. A hierarchical interaction strategy is constructed based on perturbation intensity and neighborhood information to adaptively adjust the contribution order and weighting coefficients of different data sources in the fusion process, forming a semi-serialized perturbation fusion representation. The semi-serialized perturbation fusion representation is carried out in a micro-region grid network for multiple rounds of dynamic feedback. In each iteration, the grid weights are fine-tuned according to the local perturbation trend, neighborhood influence and perturbation factor optimization results to form a fusion feature matrix, and weighted correction is performed on potential anomalies and predicted unstable areas. After the dynamic feedback converges, the cumulative effect of local fluctuations in the fused feature matrix is ​​calculated to generate a global short-term wind field disturbance distribution map, and the wind speed, wind direction and disturbance probability prediction results for the next few minutes are output.

2. The intelligent wind field forecasting method based on multi-source data fusion according to claim 1, characterized in that, The process of generating a dynamic hierarchical perturbation feature set based on local flow sensitivity and temporal continuity is as follows: By performing granular sampling of radar, satellite, ground observation station and numerical model data, a spatiotemporal index is established for various types of observation data in a micro-area multi-dimensional grid space; Based on local wind speed gradient, wind direction change rate and historical disturbance patterns, the flow sensitivity coefficient of each micro-grid point is calculated. Data is hierarchically archived according to temporal continuity and local sensitivity, generating a dynamic hierarchical perturbation feature set for cross-projection.

3. The intelligent wind field forecasting method based on multi-source data fusion according to claim 2, characterized in that, The process of cross-projecting historical short-term wind field disturbance patterns and real-time observation features onto a dynamic hierarchical disturbance feature set is as follows: The generated dynamic hierarchical perturbation feature set is used as input, and the micro-area perturbation features are aligned with historical short-term perturbation patterns according to the time window. By using feature similarity measurement methods, the mapping relationship between historical disturbance patterns and real-time observation features is identified within a dynamic hierarchical disturbance feature set, forming micro-area disturbance matching pairs; A cross-projection matrix is ​​constructed by stitching together the historical and real-time perturbation features of each micro-region grid point in the dynamic hierarchical perturbation feature set with multi-dimensional vectors.

4. The intelligent wind field forecasting method based on multi-source data fusion according to claim 3, characterized in that, The process of forming transient disturbance characteristics with adaptive response capability is as follows: A multimodal feature coding network is used to extract nonlinear features from the cross-projection matrix, and historical perturbation patterns, real-time observation features and micro-area sensitivity coefficients are embedded using tensor quantization. Based on the perturbation weight modulation mechanism, the contribution ratio of different data sources in the micro-area feature representation is adaptively adjusted, and the micro-area transient perturbation features with adaptive response capability are output.

5. The intelligent wind field forecasting method based on multi-source data fusion according to claim 4, characterized in that, The process of constructing a hierarchical interaction strategy based on perturbation strength and neighborhood information is as follows: Threshold scanning and local gradient analysis are performed on the transient disturbance characteristics of micro-regions to identify abnormal disturbance amplitudes and high-impact areas. By combining the perturbation intensity, temporal continuity, and sensitivity coefficient of outliers with those of neighboring grid points, a hierarchical interaction strategy is constructed to guide the contribution order and weight modulation of each data source.

6. The intelligent wind field forecasting method based on multi-source data fusion according to claim 5, characterized in that, The process of forming a semi-serialized perturbation fusion representation is as follows: Based on a hierarchical interaction strategy, the perturbation intensity, neighborhood influence, and micro-region sensitivity of different observation sources are comprehensively evaluated, and the contribution order of each observation source in the fusion process is dynamically arranged according to the evaluation results. Initial weighting coefficients are assigned to each observation source according to the order of contribution, and the weighting coefficients are adaptively updated according to real-time perturbation changes. Highly sensitive micro-regions are given priority to enter the fusion sequence. Short-time response weighting is applied to high-weight micro-region features in the preceding contribution position, while long-term smoothing weighting is applied to low-weight micro-region features in the following contribution position, generating a semi-serialized perturbation fusion representation.

7. The intelligent wind field forecasting method based on multi-source data fusion according to claim 6, characterized in that, The process of fusing semi-serialized perturbations into a microgrid network and performing multiple rounds of dynamic feedback is as follows: The semi-serialized perturbation fusion representation is input into the micro-grid network, and information is transferred and local features are updated according to the neighborhood influence relationship. In each round of feedback, the grid weights are adaptively fine-tuned by combining the micro-region perturbation trend and transient characteristics to form an optimized fusion representation.

8. The intelligent wind field forecasting method based on multi-source data fusion according to claim 7, characterized in that, The process of forming the fusion feature matrix is ​​as follows: The weights of micro-region grid points are adjusted based on the micro-region disturbance change trend and the differences in the characteristics of neighboring grid points in each round of dynamic feedback. By combining the optimization results of the disturbance factor, the weights of high-risk or unstable areas are enhanced or suppressed to form a fusion weight matrix, which fully reflects the cumulative effect of short-term wind field disturbance in micro-areas.

9. The intelligent wind field forecasting method based on multi-source data fusion according to claim 8, characterized in that, The process of generating a global short-term wind field disturbance distribution map is as follows: The cumulative effect of local perturbations at each micro-region grid point in the fused feature matrix is ​​calculated. By combining grid point weights, historical patterns, and neighborhood sensitivity, a comprehensive assessment and weighted correction is made for short-term wind speed, wind direction, and disturbance probability. Output the generated global short-term wind field disturbance distribution map and mark potentially high-risk areas on the map.

10. A wind field intelligent forecasting system based on multi-source data fusion, applied to the method described in any one of claims 1-9, characterized in that, include: Particle processing module: Performs granular processing on multi-source wind field data and maps it to a micro-area multi-dimensional grid space to generate a dynamic hierarchical perturbation feature set; The perturbation coding module cross-projects historical short-term wind field perturbation patterns with real-time observation features, and forms adaptive transient perturbation features through multimodal feature coding and weight modulation. Anomaly scheduling module: Scans transient disturbance features, identifies high-impact anomalies, and adaptively adjusts the fusion weights of different data sources based on disturbance intensity and neighborhood information to generate a semi-serialized disturbance fusion representation; Feedback optimization module: Performs multiple rounds of dynamic feedback on the semi-serialized fusion representation in the micro-grid network, fine-tunes the grid weights and corrects potential anomalies, and forms a fusion feature matrix; Global Prediction Module: Calculates the cumulative effect of local disturbances in the fused feature matrix, generates the global short-term wind field disturbance distribution, and outputs the wind speed, wind direction, and disturbance probability prediction for the next few minutes.

Images