Ultra-short-term wind speed prediction system combining meteorological data and multi-tower multi-layer wind measurement data

Abstract

The invention discloses an ultra-short-term wind speed prediction system combining meteorological data and multi-tower and multi-layer wind measurement data, and relates to the technical field of computer data processing, and the system comprises a self-adaptive data analysis unit, a parallel space-time processing unit, a neuromorphological feature extraction unit, a feature fusion unit and a feature distillation unit which are connected in sequence. The adaptive data analysis unit is configured to receive meteorological grid data and anemometer tower group data, and split the meteorological data into independent substreams according to a vertical height layer through a preset format topology library. According to the ultra-short-term wind speed prediction system combining the meteorological data and the multi-tower and multi-layer wind measurement data, the real-time bottleneck of high-dimensional and high-frequency meteorological and wind measurement data fusion is overcome, and the preprocessing delay of multi-source heterogeneous data is reduced; the triple verification mechanism guarantees the reliability of feature fusion, and avoids the consumption of invalid computing resources. The problem of accumulative delay caused by a traditional serial processing flow is solved, and it is ensured that a data processing link meets the timeliness constraint of ultra-short-term prediction.

Description

Technical Field

[0001] This invention relates to the field of computer data processing technology, specifically to an ultra-short-term wind speed prediction system that combines meteorological data with multi-tower, multi-layer wind measurement data. Background Technology

[0002] In wind farm operation and grid dispatching, achieving high-precision ultra-short-term wind speed forecasting is crucial. To improve forecasting performance, existing technologies tend to integrate multi-source heterogeneous data, particularly combining large-scale real-time meteorological data with dense wind measurement data from multiple meteorological towers at different heights. Theoretically, this fusion can capture more comprehensive spatiotemporal evolution information of the wind field, significantly improving the forecasting model's ability to characterize complex atmospheric dynamics. However, building and implementing such an ultra-short-term wind speed forecasting system that combines meteorological data with multi-tower, multi-layer wind measurement data faces a severe fundamental challenge. The core of the problem lies in the data processing stage: to achieve the rapid response within minute-level or even shorter time windows necessary for ultra-short-term forecasting, the system must be able to ingest, process, and analyze high-dimensional, high-frequency, multi-source heterogeneous data streams in near real-time. This data includes structurally complex gridded meteorological forecasts or observational data, as well as massive amounts of point wind measurement data collected at high frequencies from sensors at various levels of multiple meteorological towers in dispersed geographical locations.

[0003] Existing data processing architectures generally suffer from efficiency bottlenecks when handling data fusion tasks of this scale and complexity. Specifically, the data parsing and transformation process is cumbersome and time-consuming; the precise alignment and interpolation calculations of multi-source data in the spatiotemporal dimensions are computationally expensive; and the high-dimensional feature extraction required to support complex prediction models is computationally intensive. If these data processing steps are executed in a traditional serial or simple parallel manner, the accumulated latency will severely hinder the entire prediction process, making data processing itself a key bottleneck restricting the system's real-time performance. Ultimately, this results in the inability to output prediction results within an effective ultra-short-term time window, making it difficult for the system to meet the stringent timeliness requirements of practical applications. The current challenge is to efficiently and with low latency process the fusion of high-dimensional, high-frequency heterogeneous meteorological and multi-tower, multi-layer wind measurement data streams to meet the stringent real-time requirements of ultra-short-term wind speed prediction. Summary of the Invention

[0004] To achieve the above objectives, the present invention is implemented through the following technical solution: an ultra-short-term wind speed prediction system that combines meteorological data and multi-tower multi-layer wind measurement data, comprising an adaptive data parsing unit, a parallel spatiotemporal processing unit, a neuromorphic feature extraction unit, a feature fusion unit, and a feature distillation unit connected in sequence; The adaptive data parsing unit is configured to receive meteorological grid data and wind tower group data, split the meteorological data into independent sub-streams according to vertical height layers through a pre-set format topology library, and reorganize the wind measurement data into three-dimensional data blocks based on spatial topology relationships. The parallel spatiotemporal processing unit includes a meteorological channel and a wind measurement channel. The meteorological channel uses an interpolation algorithm driven by the atmospheric motion equation to process the meteorological substream, while the wind measurement channel uses a compression alignment technique driven by a turbulence correlation model to process the three-dimensional data block. The neuromorphic feature extraction unit includes a parallel convolutional spatiotemporal memory network and a graph spiking neural network, which respectively process the meteorological channel output and the wind measurement channel output; The feature fusion unit performs triple verification coupling on the two feature vectors; The feature distillation unit performs dimensionality reduction processing on the fused features to generate a prediction input vector.

[0005] Preferably, the feature fusion unit includes: The feature filtering module dynamically filters out redundant feature dimensions based on the information entropy change rate. The adversarial verification module reconstructs the feature distribution through a generator and the discriminator calculates the distribution difference between the original features and the reconstructed features. The causal reasoning module constructs a spatiotemporal causal graph model of meteorological data and wind measurement data, and verifies the physical logic of feature coupling through backpropagation gradient.

[0006] Preferably, the discriminator output difference value of the adversarial verification module is compared with a preset dynamic threshold, and feature back-transmission reprocessing is triggered when the difference value exceeds the threshold.

[0007] Preferably, the spatiotemporal causal graph of the causal reasoning module includes meteorological field pressure gradient nodes, wind measurement tower group spatial location nodes, and wind speed propagation path edges, and the rationality of feature coupling is verified by edge weight gradient.

[0008] Preferably, it also includes a two-stage verification loop: In the first verification phase, a real-time evaluator is set up after the feature fusion unit to monitor changes in feature vector entropy and trigger reprocessing instructions. In the second verification stage, a simulation predictor is set up after the feature distillation unit to back-map the dimensionality-reduced features to the original data space to calculate the reconstruction error.

[0009] Preferably, the real-time evaluator is configured to send a reprocessing instruction to the neuromorphic feature extraction unit when the rate of change of the feature vector entropy value exceeds a set range.

[0010] Preferably, the meteorological channel and the wind measurement channel of the parallel spatiotemporal processing unit share a memory resource pool, and the processing resources are allocated through dynamic scheduling of computing tasks.

[0011] Preferably, the convolutional spatiotemporal memory network includes spatiotemporal convolutional layers and gated memory units, used to extract multi-scale motion patterns from meteorological data.

[0012] Preferably, the graph pulse neural network includes a spatial graph convolutional layer and a temporal pulse encoder to capture the dynamic correlation features among the anemometer tower groups.

[0013] According to an ultra-short-term wind speed prediction system that combines meteorological data with multi-tower, multi-layer wind measurement data, the prediction input vector is input to the wind speed prediction model to output the wind speed prediction result.

[0014] This invention provides an ultra-short-term wind speed prediction system that combines meteorological data with multi-tower, multi-layer wind measurement data. It has the following beneficial effects: This ultra-short-term wind speed forecasting system, which combines meteorological data with multi-tower, multi-layer anemometer data, effectively overcomes the real-time bottleneck of fusing high-dimensional, high-frequency meteorological and anemometer data by constructing a streaming heterogeneous data collaborative processing engine and employing a spatiotemporally decoupled dual-channel architecture and a neuromorphic feature extraction mechanism. An adaptive data parsing matrix dynamically optimizes the data splitting path, and a differentiated acceleration strategy is implemented in conjunction with a parallel spatiotemporal preprocessor to reduce the preprocessing latency of multi-source heterogeneous data. A triple verification mechanism ensures the reliability of feature fusion and avoids unnecessary computational resource consumption. It solves the cumulative latency problem caused by traditional serial processing flows, ensuring that the data processing stage meets the timeliness constraints of ultra-short-term forecasting.

[0015] This ultra-short-term wind speed forecasting system, which combines meteorological data with multi-tower, multi-layer wind measurement data, achieves the co-evolution of data processing and forecasting decisions based on a closed-loop feedback dynamic optimization system: a two-stage validation loop monitors feature quality and output error in real time, maintaining stable system operation through reprocessing instructions with dimensional indexes and a backup model switching mechanism; a spatial coordinate-meteorological condition correlation graph accurately locates the source of forecasting error, driving the data parsing unit to optimize the vertical layering strategy and the feature extraction unit to adjust channel sensitivity; confidence vector fusion of multi-source validation information guides the dynamic adjustment of the forecasting cycle, and a high-frequency data acquisition mechanism rapidly responds to meteorological changes. It breaks through the limitations of traditional systems where "data processing" and "forecasting application" are separate, continuously improving forecast accuracy and system fault tolerance in complex meteorological scenarios. Attached Figure Description

[0016] Figure 1 This is a schematic diagram of the module interaction of an ultra-short-term wind speed prediction system that combines meteorological data with multi-tower, multi-layer wind measurement data according to the present invention. Figure 2 This is a flowchart illustrating an ultra-short-term wind speed prediction method that combines meteorological data with multi-tower, multi-layer wind measurement data according to the present invention. Detailed Implementation

[0017] 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.

[0018] Please see Figure 1 and Figure 2 The present invention provides a technical solution: an ultra-short-term wind speed prediction system that combines meteorological data and multi-tower multi-layer wind measurement data, comprising an adaptive data parsing unit, a parallel spatiotemporal processing unit, a neuromorphic feature extraction unit, a feature fusion unit and a feature distillation unit connected in sequence; The adaptive data parsing unit is configured to receive meteorological grid data and wind tower group data, and split the meteorological data into independent sub-streams according to vertical height layers through a pre-set format topology library. At the same time, it reassembles the wind measurement data into three-dimensional data blocks based on spatial topology relationships. The parallel spatiotemporal processing unit includes a meteorological channel and a wind measurement channel. The meteorological channel uses an interpolation algorithm driven by the atmospheric motion equation to process the meteorological substream, while the wind measurement channel uses a compression alignment technique driven by a turbulence correlation model to process the three-dimensional data block. The neuromorphic feature extraction unit comprises a parallel convolutional spatiotemporal memory network and a graph spiking neural network, which process the meteorological channel output and the wind measurement channel output, respectively. The feature fusion unit performs triple verification coupling on the two feature vectors; The feature distillation unit performs dimensionality reduction on the fused features to generate the prediction input vector.

[0019] It should be further explained that, in the specific implementation process, upon system startup, the adaptive data parsing unit receives externally input meteorological grid data and raw data collected by the distributed anemometer tower group. The meteorological grid data identifies its three-dimensional structural features through a pre-defined format topology library and decouples it into independent meteorological substreams according to vertical height layers; simultaneously, the anemometer tower group data constructs spatial topological relationships based on geographic coordinates and reorganizes into three-dimensional data blocks containing tower locations and height dimensions. The two types of data are input into the spatiotemporal processing unit in parallel: the meteorological substream enters the meteorological channel and is processed using an interpolation algorithm based on atmospheric motion equations. This algorithm dynamically adjusts the interpolation weights by combining the relationship between pressure gradient and temperature field changes; the anemometer data blocks are input into the anemometer channel and a compression alignment technique driven by a turbulence correlation model is applied to compress redundant information by establishing vortex transfer functions between adjacent tower layers.

[0020] The processed dual-channel data are input into the neuromorphic feature extraction unit. The meteorological substream is processed by a convolutional spatiotemporal memory network, whose spatiotemporal convolutional layers extract atmospheric motion patterns at different scales, and the gated memory unit tracks the migration trajectory of the pressure system. The wind measurement data block is processed by a graph spiking neural network, where the spatial graph convolutional layer associates the positional relationships of the wind measurement towers, and the temporal pulse encoder captures the phase propagation characteristics of turbulent energy. The feature vectors output from both channels are input into the feature fusion unit to perform triple verification coupling: first, feature dimensions with contributions below a threshold are dynamically filtered out based on the rate of change of information entropy; then, the adversarial verification module reconstructs the feature probability distribution through a generator, and the discriminator calculates the KL divergence between the original features and the reconstructed features. When the divergence value exceeds the dynamically adjusted tolerance range, feature backpropagation is triggered; finally, the causal reasoning module constructs a directed graph of the meteorological field pressure nodes and the wind measurement tower position nodes, calculates the gradient loss in reverse along the wind speed propagation path, and initiates feature reorganization if the loss value indicates a logical conflict.

[0021] The fused feature vector undergoes hierarchical compression via a feature distillation unit: the first-layer attention mechanism focuses on high-information-density dimensions, while the second-layer sparse coding eliminates linear correlations between features, generating a low-dimensional prediction input vector. Throughout the processing, a two-stage validation loop continuously monitors data quality: when the entropy change rate output by the feature fusion unit exceeds the normal fluctuation range, the real-time evaluator sends a reprocessing instruction to the feature extraction unit; the prediction input vector output by the feature distillation unit is back-mapped to the original data space by the simulated predictor, and if the reconstruction error continues to increase, a system self-check is initiated.

[0022] The feature fusion unit includes: The feature filtering module dynamically filters out redundant feature dimensions based on the information entropy change rate. The adversarial verification module reconstructs the feature distribution through a generator and the discriminator calculates the distribution difference between the original features and the reconstructed features. The causal reasoning module constructs a spatiotemporal causal graph model of meteorological data and wind measurement data, and verifies the physical logic of feature coupling through backpropagation gradient.

[0023] It should be further explained that, in the specific implementation process, after the feature fusion unit receives the meteorological feature vector and the wind measurement feature vector from the neuromorphic feature extraction unit, it initiates a triple verification coupling process. First, in the feature screening module, the information entropy change rate of each dimension of the meteorological feature vector is calculated. When the entropy change rate of a specific dimension is continuously lower than a set threshold, it is determined that the dimension does not contribute enough to the dynamic representation of the wind field and is removed from the feature set to be fused. At the same time, the fluctuation range of the entropy value of the wind measurement feature vector is monitored, and a manual review mechanism is initiated for dimensions with abnormally narrowed fluctuation ranges.

[0024] The filtered feature set is input into the adversarial verification module: the generator generates simulated wind measurement features based on the distribution of meteorological features, and the discriminator compares the KL divergence values ​​of the real wind measurement features and the generated features; the system presets a dynamic tolerance range, and when the KL divergence value falls into the lower limit of the range, it is determined that the feature information is redundant and triggers the feature dimensionality reduction and reprocessing instruction; when the divergence value exceeds the upper limit of the range, it is determined that the feature is mismatched and sends a data re-acquisition request to the data parsing unit.

[0025] Finally, a directed acyclic graph (DAG) structure is constructed using a causal reasoning module. This graph includes isobar nodes for the meteorological field, geographical location nodes for the anemometer towers, and wind speed propagation path edges derived from atmospheric dynamics. During backpropagation calculations, if the sign correlation between the output layer's predicted wind speed gradient and the input layer's meteorological pressure gradient is contradictory, it is determined that the feature coupling violates physical laws. The conflicting feature dimensions are automatically isolated, and alternative feature combination schemes are initiated. Throughout the entire process, any abnormal signal output at any verification stage will interrupt subsequent fusion operations until the self-check of that stage is completed.

[0026] The discriminator outputs a difference value in the adversarial verification module and compares it with a preset dynamic threshold. When the difference value exceeds the threshold, feature reprocessing is triggered. It should be further explained that, in the specific implementation, when the adversarial verification module is running, the generator receives meteorological feature vectors as input and reconstructs the probability distribution of wind measurement features through a deep probabilistic generation network to generate simulated wind measurement feature vectors. The discriminator simultaneously acquires the real wind measurement feature vectors and calculates the KL divergence values ​​of the simulated and real features in the temporal spectral feature dimension and the spatial correlation dimension, respectively. The system dynamically adjusts the tolerance interval based on the feature stability within the historical data window: compressing the interval width when the standard deviation of the feature distribution narrows for multiple consecutive periods, and expanding the upper limit of the interval when a sudden meteorological disturbance is detected.

[0027] The discriminator compares the KL divergence value with the current tolerance range, triggering a tiered response mechanism. This includes: if the divergence value is within the tolerance range, marking the simulated feature as valid and outputting it to the next stage; if the divergence value is below the lower limit of the range, determining feature information redundancy, sending a dimensionality reduction instruction to the feature distillation unit, and freezing the current generator parameters; if the divergence value exceeds the upper limit of the range, immediately interrupting the fusion process and sending a re-acquisition request code to the data parsing unit. This code contains an abnormal feature dimension identifier to guide the data source to perform targeted re-acquisition. When the upper limit alarm is triggered continuously, the system automatically switches to a backup generator model and resets the discriminator weights, while simultaneously initiating a feature channel self-diagnosis program to verify the integrity of data transmission. All anomaly handling records are written to the verification log for optimizing the tolerance range adjustment strategy.

[0028] The spatiotemporal causal graph of the causal inference module includes meteorological field pressure gradient nodes, anemometer cluster spatial location nodes, and wind speed propagation path edges. The rationality of feature coupling is verified through edge weight gradients. It should be further explained that, in the specific implementation process, when the causal inference module starts, it constructs pressure gradient nodes based on the curvature distribution of isobars in the current meteorological field, with node weights correlated to the rate of change of the distance between adjacent isobars. Simultaneously, it generates spatial location nodes based on the geographical coordinates of the anemometer cluster, with node attributes including altitude and surrounding terrain shading coefficients. Wind speed propagation path edges are created between nodes, with the edge direction pointing from the pressure gradient node to the location node. The initial weights are set based on the relationship between pressure gradient force and geostrophic wind speed in atmospheric dynamics.

[0029] During feature coupling verification, bidirectional gradient calculations are performed along the propagation path edges: forward propagation maps meteorological features to predicted wind speed distribution, and backward propagation calculates the partial derivatives of wind speed changes at location nodes with respect to pressure gradient nodes. When the sign of the backward gradient at a specific path edge conflicts with the physical laws of the pressure gradient, including situations such as a negative gradient at a high-pressure area location node but a positive feedback weight is detected, the path logic is deemed abnormal.

[0030] The system automatically isolates the feature dimensions associated with conflicting edges and activates alternative propagation path networks. These alternative networks utilize path combinations validated under historical conditions, reconstructing the inference graph by attenuating the weights of anomalous edges and reinforcing statistically significant paths. Pressure gradient nodes that continuously trigger conflict determinations are marked as untrusted nodes and require re-verification in the next data processing cycle. The weight gradients of all path edges are stored in a causal knowledge base for optimizing the initial weight setting strategy.

[0031] It also includes a two-stage verification loop: In the first verification phase, a real-time evaluator is set up after the feature fusion unit to monitor changes in feature vector entropy and trigger reprocessing instructions. In the second verification stage, a simulation predictor is set up after the feature distillation unit to back-map the dimensionality-reduced features to the original data space to calculate the reconstruction error.

[0032] It should be further explained that, in the specific implementation process, the two-stage verification loop is activated synchronously during system operation. The first verification stage is deployed at the output of the feature fusion unit. The real-time evaluator continuously monitors the information entropy value of the fused feature vector. When it detects that the rate of change of the entropy value exceeds the historical normal fluctuation range, including: for example, the rate of change increases for three consecutive sampling periods and deviates from the baseline standard deviation by more than twice, the following steps are taken: the feature quality is judged to be abnormal, and a reprocessing instruction code is immediately sent to the neuromorphic feature extraction unit. This code contains an identifier for the abnormal entropy change dimension. After receiving the instruction, the feature extraction unit first restarts the feature extraction channel pointed to by the identifier, and retains the processing results of the unaffected dimensions.

[0033] The second verification stage is located after the feature distillation unit. The simulation predictor reconstructs the original data space representation from the dimensionality-reduced prediction input vector. It calculates the grid root mean square deviation between the reconstructed meteorological data and the original meteorological data, and the tower correlation coefficient between the reconstructed wind measurement data and the original wind measurement data. If the meteorological deviation continues to increase and the wind measurement correlation coefficient continues to decrease, the system self-check protocol is activated: first, the output port of the feature distillation unit is isolated, and a backup dimensionality reduction model is used to generate a temporary prediction input vector; at the same time, the data parsing unit is triggered to perform timestamp integrity verification on the original data to eliminate time domain misalignment caused by data transmission delay.

[0034] The prediction results generated during the self-inspection period are marked with a confidence level indicator, and the isolation status is lifted after the deviation and correlation coefficient return to the normal threshold. The two-stage verification logs are synchronized to the system console in real time to form a closed-loop diagnostic report.

[0035] The real-time evaluator is configured to send a reprocessing instruction to the neuromorphic feature extraction unit when the rate of change of the feature vector entropy exceeds a set range. It should be further noted that, in practice, the real-time evaluator continuously tracks the rate of change of the feature vector entropy output by the feature fusion unit during operation, maintaining a historical normal fluctuation benchmark through a dynamic learning module. When the entropy rate of a specific dimension continuously exceeds the benchmark range, a multi-level judgment process is initiated: first, it checks whether the abnormal dimension is spatiotemporally correlated with recent meteorological disturbances; if a correlation exists, the judgment threshold is relaxed by one level; if it is an isolated anomaly, the information contribution weight of that dimension in the feature vector is further analyzed, and dimensions with weights below the critical value are only marked and not triggered with instructions.

[0036] When intervention is necessary, a reprocessing instruction code containing an anomaly dimension index is generated and sent to the neuromorphic feature extraction unit, prioritizing non-peak periods. Upon receiving the instruction, the feature extraction unit executes a step-by-step response: only the feature extraction channel pointed to by the index is restarted, while the processing results of the other channels are retained; the restart process adopts a progressive initialization strategy, first loading the historical feature average of the previous three periods of the anomaly dimension as the initial state, and then gradually fusing the real-time data stream.

[0037] If reprocessing is triggered consecutively for the same dimension, the system automatically increases the priority of that channel in the computing resource pool and generates a dimension health report, which is then pushed to the operations and maintenance interface. All reprocessing operations record dimension index and response latency data to optimize command sending strategies.

[0038] The meteorological and anemometer channels of the parallel spatiotemporal processing unit share a memory resource pool, and resource allocation is achieved through dynamic scheduling of computational tasks. It should be further explained that, in the specific implementation, during the operation of the parallel spatiotemporal processing unit, the meteorological and anemometer channels coordinate the allocation of computational resources through the shared memory resource pool. During resource pool initialization, basic computing units are reserved for the minimum guarantee of use by both channels, and the remaining resources form a dynamic allocation pool. The resource scheduler continuously monitors the processing status of both channels: when the meteorological channel detects a sudden change in air pressure gradient, it sends a resource preemption request code to the scheduler, which includes the required number of computing units and the expected duration; after verifying the rationality of the request, the scheduler temporarily allocates resources from the dynamic pool to the meteorological channel, and simultaneously initiates a buffer queue mechanism for the anemometer channel to temporarily store data blocks to be processed.

[0039] If the wind measurement channel triggers a load alarm due to a surge in tower data, the scheduler adopts a gradual resource release strategy, including: prioritizing the reduction of resource allocation for non-core interpolation operations in the meteorological channel, and transferring computing units to the wind measurement channel in batches.

[0040] The resource transfer process implements state synchronization: the meteorological channel freezes the current processing frame to save the intermediate state, and the wind measurement channel loads historical parameters from the same operating condition model to preheat the newly allocated resources. When the dual-channel load returns to the balanced range, the scheduler initiates a resource reclamation protocol: gradually reclamating temporary allocation units and releasing the wind measurement buffer queue, during which the timeliness of data block processing is verified, and accelerated processing threads are started for data blocks nearing timeout. All resource scheduling records generate optimization logs, which are used to train the load prediction model to improve scheduling foresight.

[0041] The convolutional spatiotemporal memory network comprises spatiotemporal convolutional layers and gated memory units, used to extract multi-scale motion patterns from meteorological data. It should be further noted that, in practical implementation, when processing meteorological substreams, the spatiotemporal convolutional layers employ a multi-scale sliding window sampling mechanism. When inputting pressure field data, the direction of regional pressure gradient changes is first detected, and elongated convolutional kernels are deployed along the gradient direction to capture the migration characteristics of the pressure system; for temperature field data, ring-shaped convolutional kernels are used to identify the evolution patterns of thermal centers.

[0042] The gated memory unit dynamically tracks the evolution trajectory of key meteorological elements: during the initialization phase, regions with continuously increasing pressure gradients are marked as focal regions, and the memory unit is assigned additional storage weights; when the rate of change of pressure gradient in the focal region exceeds the historical benchmark, the convolution kernel reorganization mechanism is triggered, including: shrinking the kernel size in the low-change direction and expanding the receptive field in the high-change direction.

[0043] The memory unit periodically evaluates its storage status, reducing the memory weights of weather models that have been inactive for three consecutive cycles, freeing up space for newly formed perturbation systems. When handling sudden weather events, the network automatically switches to a backup memory channel: this channel pre-stores typical weather process models, such as frontal processes, and quickly loads corresponding parameters through similarity matching. The weight update records of all convolutional kernels are input into the feature analysis module to optimize the initial convolutional kernel configuration strategy.

[0044] The graph spiking neural network (PSN) comprises a spatial graph convolutional layer and a temporal pulse encoder to capture the dynamic correlation features among anemometer towers. It should be further noted that, in the specific implementation, during PSN initialization, a spatial graph structure is constructed based on the three-dimensional coordinates of the anemometer towers, and the connection weights between nodes are calibrated using both terrain shading effects and historical turbulence correlation strength. When processing anemometer data blocks, the spatial graph convolutional layer transmits turbulence energy characteristics along the connection edges, and a flow compensation coefficient is added to tower nodes in valley terrain areas.

[0045] The time-pulse encoder employs a phase modulation mechanism: when the prevailing wind direction is detected to be consistently stable, the feature propagation phase is locked to reduce computational overhead; when a sudden change in wind direction exceeds a threshold angle, multi-phase parallel encoding is initiated to capture transient turbulence features. The network dynamically monitors the energy transfer efficiency between nodes, deweights connection edges with continuously increasing transmission delays, and activates alternative topology paths, including connection patterns that have been validated under historically similar wind direction conditions.

[0046] When processing data from strongly sheared layers, a vertical momentum correction factor is automatically injected to suppress feature distortion. The impulse activation modes of all nodes are recorded in the feature evolution map to optimize the initial connection weight allocation strategy.

[0047] According to an ultra-short-term wind speed prediction system that combines meteorological data with multi-tower, multi-layer wind measurement data, the prediction input vector is fed into the wind speed prediction model to output the wind speed prediction result. It should be further noted that, in the specific implementation process, when the prediction input vector is input into the wind speed prediction model, the model prioritizes loading a parameter set that matches the current atmospheric stability level. The processing continuously monitors the activity index of each dimension in the input vector. When topographic disturbance-related dimensions are detected to continuously dominate the feature space, the decision weight of the local topographic correction submodule is automatically increased; if the activity of the thermal convection dimension suddenly increases, the boundary layer parameterization scheme is activated to suppress excessive response.

[0048] While outputting the prediction results, the model generates a confidence assessment vector, which integrates three pieces of information: the triple validation status code of the feature fusion unit, the anomaly marker of the two-stage validation loop, and the residual analysis results of the prediction model itself. When the confidence assessment indicates that the meteorological system is in a rapid evolution phase, the system automatically shortens the time window of the next prediction cycle and starts a high-frequency data acquisition mode to supplement the observation blind spots.

[0049] After comparing all predicted results with measured data, an error distribution map is generated. The map marks the spatial coordinates and corresponding meteorological conditions of areas with persistently high errors. This information is fed back in real time to the adaptive data parsing unit to optimize the vertical layering strategy, and simultaneously guides the neuromorphic feature extraction unit to adjust the sensitivity parameters of the feature extraction channels. Intermediate data from the entire prediction process is stored in a rolling buffer, supporting the reconstruction of the feature processing state at any point in time during backtracking analysis.

[0050] It should be further explained that, in the specific implementation process, the adaptive data parsing unit receives meteorological grid data and raw data from the anemometer tower group upon system startup. The meteorological data is analyzed using a pre-defined format parsing library to identify its three-dimensional structural features, and then split into independent substreams according to vertical height layers. The splitting process dynamically adjusts the interlayer density based on the atmospheric boundary layer height. The anemometer tower group data constructs a spatial topology network based on geographic coordinates, reorganizing the discrete tower layer observations into three-dimensional data blocks with spatial location identifiers. Both types of data are input into the spatiotemporal processing unit in parallel: the meteorological substream enters the meteorological processing channel, where an interpolation algorithm based on atmospheric motion equations is used. This algorithm dynamically corrects the interpolation weights by combining the pressure gradient and temperature advection relationship; the anemometer data blocks are input into the anemometer processing channel, where a compression alignment technique driven by a turbulence correlation model is applied, eliminating data redundancy by establishing vortex transfer functions between adjacent tower layers.

[0051] The processed dual-channel data enters the neuromorphic feature extraction stage. The meteorological substream is processed by a convolutional spatiotemporal memory network, whose multi-scale convolutional kernels capture the movement features of the weather system along the pressure gradient direction, and the gated memory unit tracks the evolution trajectory of meteorological elements in highly variable areas. The wind measurement data block is processed by a graph spiking neural network, and the spatial graph convolutional layer integrates the terrain shading effect to calculate the node correlation. The temporal spiking encoder switches the encoding phase according to the stability of the wind direction. The feature vectors output from the two channels are input into the feature fusion unit to perform triple verification: First, feature dimensions with insufficient contribution are dynamically screened based on the information entropy change rate, and the screening process refers to the feature validity records under the same historical conditions; then, the adversarial verification module reconstructs the feature probability distribution through the generator, and the discriminator calculates the distribution difference value between the original feature and the reconstructed feature. When the difference value exceeds the dynamic tolerance range, the feature backpropagation mechanism is triggered; finally, the causal reasoning module constructs a directed graph model of meteorological pressure nodes and wind measurement location nodes, verifies the gradient sign consistency along the wind speed propagation path, and isolates conflicting features when physical contradictions such as negative wind speed feedback in high-pressure areas are detected.

[0052] The fused features are compressed into low-dimensional vectors via a distributed distillation pipeline. This process employs an attention mechanism to focus on high-information dimensions and eliminates linear correlations between features through sparse coding. The distilled prediction input vector is then fed into the wind speed prediction model. The model loads the corresponding parameter set based on the current atmospheric stability and outputs a prediction result, simultaneously generating a confidence vector. The confidence vector incorporates three pieces of information: feature fusion verification status identifier, data processing anomaly marker, and model residual analysis conclusion. When the confidence indicates rapid evolution of the meteorological system, the system automatically shortens the prediction period and initiates high-frequency data acquisition.

[0053] The entire process is monitored by a two-stage verification loop: In the first stage, a real-time evaluator is set up after feature fusion. When the entropy change rate of the feature vector exceeds the historical normal fluctuation range, a reprocessing instruction with dimension index is sent to the feature extraction unit. In the second stage, after the predicted input vector is output, it is back-mapped to the original data space through a simulation predictor. The correlation coefficient between the meteorological grid reconstruction deviation and the wind measurement data is calculated. If anomalies persist, a backup dimensionality reduction model is activated, and the integrity of data transmission is verified. All intermediate processing data is stored in a rolling buffer to support state backtracking. The spatial coordinate-meteorological condition correlation graph generated by the prediction error analysis is fed back to the front-end unit, guiding the data parsing unit to optimize the vertical layering strategy and the feature extraction unit to adjust the channel sensitivity.

[0054] A method for ultra-short-term wind speed prediction that combines meteorological data with multi-tower, multi-level wind measurement data includes the following steps: Step S1: The adaptive data parsing unit receives meteorological grid data and raw data from the wind measurement tower group, splits the meteorological data into independent substreams according to vertical height layers, and reassembles the wind measurement data into three-dimensional blocks based on spatial topology. Step S2: The parallel spatiotemporal processing unit starts dual-channel processing: the meteorological channel adopts a dynamic interpolation algorithm driven by the atmospheric motion equation, and the wind measurement channel applies a compression alignment technique driven by the turbulence correlation model; Step S3: The neuromorphic feature extraction unit simultaneously processes dual-channel data: the convolutional spatiotemporal memory network extracts meteorological multi-scale motion features, and the graph spiking neural network captures the dynamic correlation features of the wind tower group; Step S4: The feature fusion unit performs triple verification coupling: based on the information entropy change rate, low contribution feature dimensions are screened out; the adversarial verification module reconstructs the feature distribution through the generator; the discriminator compares the distribution difference values ​​and triggers a dynamic tolerance response; the causal reasoning module constructs a directed graph of meteorological pressure nodes and wind measurement location nodes to verify the physical consistency of gradient signs. Step S5: Distributed feature distillation pipeline compresses and fuses features: an attention mechanism is used to focus on high-information dimensions, and sparse coding eliminates linear correlations between features; Step S6: Load the atmospheric stability matching parameter set into the wind speed prediction model, and simultaneously generate a confidence vector of the fused triple verification state from the output prediction results; Step S7: Real-time monitoring of the two-stage verification loop: After feature fusion, the real-time evaluator monitors the entropy change rate and triggers a reprocessing instruction with dimension index when an anomaly occurs; after the predicted input vector is output, the simulation predictor back-maps and calculates the reconstruction bias and correlation coefficient, and starts the backup dimensionality reduction model when an anomaly continues. Step S8: Prediction error analysis generates a spatial coordinate-meteorological condition correlation map, which is fed back to the data parsing unit to optimize the vertical layering strategy and adjust the channel sensitivity of the feature extraction unit; Step S9: The rolling buffer stores intermediate data throughout the entire process, supporting backtracking of the processing status of any node; Step S10: Dynamically adjust the prediction period based on the confidence vector: shorten the period and activate high-frequency data acquisition when the meteorological system evolves rapidly.

[0055] By constructing a streaming heterogeneous data collaborative processing engine, and employing a spatiotemporally decoupled dual-channel architecture and a neuromorphic feature extraction mechanism, the real-time bottleneck of high-dimensional, high-frequency meteorological and wind measurement data fusion is effectively overcome. An adaptive data parsing matrix dynamically optimizes the data splitting path, and a differentiated acceleration strategy is implemented in conjunction with a parallel spatiotemporal preprocessor to reduce the preprocessing latency of multi-source heterogeneous data. A triple verification mechanism ensures the reliability of feature fusion and avoids unnecessary computational resource consumption. This addresses the cumulative latency problem caused by traditional serial processing flows, ensuring that the data processing stage meets the timeliness constraints of ultra-short-term forecasting.

[0056] A dynamic optimization system based on closed-loop feedback enables the co-evolution of data processing and predictive decision-making: a two-stage validation loop monitors feature quality and output error in real time, maintaining stable system operation through reprocessing instructions with dimensional indexes and a backup model switching mechanism; a spatial coordinate-meteorological condition correlation graph accurately locates the source of prediction error, driving the data parsing unit to optimize the vertical layering strategy and the feature extraction unit to adjust channel sensitivity; confidence vector fusion of multi-source validation information guides the dynamic adjustment of the prediction cycle, and combined with a high-frequency data acquisition mechanism, it quickly responds to meteorological changes. This system overcomes the limitations of the traditional system's separation of "data processing" and "predictive application," continuously improving prediction accuracy and system fault tolerance in complex meteorological scenarios.

[0057] 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 a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0058] 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 short-term wind speed prediction system that combines meteorological data with multi-tower, multi-layer wind measurement data, characterized in that, It includes an adaptive data parsing unit, a parallel spatiotemporal processing unit, a neuromorphic feature extraction unit, a feature fusion unit, and a feature distillation unit connected in sequence; The adaptive data parsing unit is configured to receive meteorological grid data and wind tower group data, split the meteorological data into independent sub-streams according to vertical height layers through a pre-set format topology library, and reorganize the wind measurement data into three-dimensional data blocks based on spatial topology relationships. The parallel spatiotemporal processing unit includes a meteorological channel and a wind measurement channel. The meteorological channel uses an interpolation algorithm driven by the atmospheric motion equation to process the meteorological substream, while the wind measurement channel uses a compression alignment technique driven by a turbulence correlation model to process the three-dimensional data block. The neuromorphic feature extraction unit includes a parallel convolutional spatiotemporal memory network and a graph spiking neural network, which respectively process the meteorological channel output and the wind measurement channel output; The feature fusion unit performs triple verification coupling on the two feature vectors; The feature distillation unit performs dimensionality reduction processing on the fused features to generate a prediction input vector.

2. The ultra-short-term wind speed prediction system combining meteorological data and multi-tower, multi-layer wind measurement data as described in claim 1, characterized in that: The feature fusion unit includes: The feature filtering module dynamically filters out redundant feature dimensions based on the information entropy change rate. The adversarial verification module reconstructs the feature distribution through a generator and uses a discriminator to calculate the distribution difference between the original features and the reconstructed features. The causal reasoning module constructs a spatiotemporal causal graph model of meteorological data and wind measurement data, and verifies the physical logic of feature coupling through backpropagation gradient.

3. The ultra-short-term wind speed prediction system combining meteorological data and multi-tower, multi-layer wind measurement data according to claim 2, characterized in that: The discriminator output difference value of the adversarial verification module is compared with a preset dynamic threshold. When the difference value exceeds the threshold, feature retransmission and reprocessing are triggered.

4. The ultra-short-term wind speed prediction system combining meteorological data and multi-tower, multi-layer wind measurement data according to claim 2, characterized in that: The spatiotemporal causal graph of the causal reasoning module includes meteorological field pressure gradient nodes, anemometer tower group spatial location nodes, and wind speed propagation path edges. The rationality of feature coupling is verified by edge weight gradient.

5. The ultra-short-term wind speed prediction system combining meteorological data and multi-tower, multi-layer wind measurement data according to claim 1, characterized in that: It also includes a two-stage verification loop: In the first verification phase, a real-time evaluator is set up after the feature fusion unit to monitor changes in feature vector entropy and trigger reprocessing instructions. In the second verification stage, a simulation predictor is set up after the feature distillation unit to back-map the dimensionality-reduced features to the original data space to calculate the reconstruction error.

6. The ultra-short-term wind speed prediction system combining meteorological data and multi-tower, multi-layer wind measurement data according to claim 5, characterized in that: The real-time evaluator is configured to send a reprocessing instruction to the neuromorphic feature extraction unit when the rate of change of the feature vector entropy value exceeds a set range.

7. The ultra-short-term wind speed prediction system combining meteorological data and multi-tower, multi-layer wind measurement data according to claim 1, characterized in that: The meteorological channel and wind measurement channel of the parallel spatiotemporal processing unit share a memory resource pool, and the processing resources are allocated through dynamic scheduling of computing tasks.

8. The ultra-short-term wind speed prediction system combining meteorological data and multi-tower, multi-layer wind measurement data according to claim 1, characterized in that: The convolutional spatiotemporal memory network includes spatiotemporal convolutional layers and gated memory units, used to extract multi-scale motion patterns from meteorological data.

9. The ultra-short-term wind speed prediction system combining meteorological data and multi-tower, multi-layer wind measurement data according to claim 1, characterized in that: The graph pulse neural network includes a spatial graph convolutional layer and a temporal pulse encoder, used to capture the dynamic correlation features among the wind measurement tower groups.

10. The system according to any one of claims 1-9, characterized in that, The prediction input vector is input to the wind speed prediction model to output the wind speed prediction result.

Images