An ultra-short-term wind speed prediction system combining meteorological data with multi-tower multi-layer wind measurement data

Abstract

The application discloses a kind of super-short-term wind speed prediction systems of meteorological data and multi-tower multi-layer wind data combination, the present application relates to computer data processing technical field, including sequentially connected adaptive data analysis unit, parallel space-time processing unit, neural morphological feature extraction unit, feature fusion unit and feature distillation unit;The adaptive data analysis unit is configured to receive meteorological grid data and wind tower group data, meteorological data is split into independent sub-flow according to vertical height layer by preformat topology library.The super-short-term wind speed prediction system of meteorological data and multi-tower multi-layer wind data combination, overcome the real-time bottleneck of high-dimensional high-frequency meteorological and wind data fusion, reduce the pre-processing delay of multi-source heterogeneous data;Triple verification mechanism guarantees feature fusion reliability, avoid invalid computing resource consumption.Solve the cumulative delay problem caused by traditional serial processing process, ensure that data processing link meets the timeliness constraint of super-short-term prediction.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of computer data processing, in particular to a kind of super-short-term wind speed prediction system combining meteorological data and multi-tower multi-layer wind data. BACKGROUND

[0002] In the field of wind farm operation and power grid dispatching, it is crucial to achieve high-precision super-short-term wind speed prediction. To improve prediction performance, existing technical solutions tend to integrate multi-source heterogeneous data, especially combining large-scale real-time meteorological data with intensive wind data from different height layers of multiple wind towers. This integration can theoretically capture more comprehensive spatiotemporal evolution information of wind fields, significantly improving the ability of prediction models to depict complex atmospheric dynamics. However, there is a serious fundamental challenge in constructing and implementing such a super-short-term wind speed prediction system that combines meteorological data and multi-tower multi-layer wind data. The core of the problem lies in the data processing link: in order to achieve the rapid response within the minute-level or even shorter time window necessary for super-short-term prediction, the system must be able to ingest, process, and analyze high-dimensional, high-frequency, and multi-source heterogeneous data streams in near real-time. These data include complex structured gridded meteorological forecasts or observations, as well as massive point wind data collected at high frequency from sensors at various layers of multiple wind towers in dispersed geographical locations.

[0003] The data processing architecture of existing technologies generally has efficiency bottlenecks when dealing with such large-scale and complex data integration tasks. Specifically, the data parsing and conversion process is time-consuming and cumbersome, the precise alignment and interpolation calculation of multi-source data in the time and space dimensions is computationally intensive, and the high-dimensional feature extraction operations required to support complex prediction models are also computationally intensive. If these data processing steps are executed in a traditional serial or simple parallel manner, the cumulative delay will seriously hinder the entire prediction process, making data processing itself a critical shortcoming that restricts the real-time performance of the system. Ultimately, the prediction results cannot be output in time within the effective super-short-term time window, and the system cannot meet the stringent requirements of prediction timeliness in practical applications. The current problem to be solved is: how to efficiently and low-latency process the integration of high-dimensional and high-frequency heterogeneous meteorological and multi-tower multi-layer wind data streams to meet the stringent real-time requirements of super-short-term wind speed prediction. SUMMARY

[0004] To achieve the above purpose, the present application realizes the following technical solutions: a super-short-term wind speed prediction system combining meteorological data and multi-tower multi-layer wind 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.

[0005] The adaptive data parsing unit is configured to receive meteorological grid data and wind measurement tower group data, split the meteorological data into independent sub-streams according to vertical height layers through a preset format topology library, and recombine the wind measurement data into three-dimensional data blocks based on spatial topological relations;

[0006] The parallel space-time processing unit includes a meteorological channel and a wind measurement channel, the meteorological channel processes the meteorological sub-streams by using an interpolation algorithm driven by an atmospheric motion equation, and the wind measurement channel processes the three-dimensional data blocks by using a compression alignment technology driven by a turbulence correlation model;

[0007] The neuromorphic feature extraction unit includes a convolutional space-time memory network and a graph spiking neural network connected in parallel, which respectively process the output of the meteorological channel and the output of the wind measurement channel;

[0008] The feature fusion unit performs triple verification coupling on the two feature vectors;

[0009] The feature distillation unit performs dimension reduction processing on the fused features to generate a prediction input vector.

[0010] Preferably, the feature fusion unit includes:

[0011] A feature screening module that dynamically screens out redundant feature dimensions based on information entropy change rate;

[0012] An adversarial verification module that reconstructs feature distribution by a generator and calculates the distribution difference value of the original feature and the reconstructed feature by a discriminator;

[0013] A causal reasoning module that constructs a space-time causal graph model of meteorological data and wind measurement data, and verifies the physical logic of feature coupling by backpropagation gradient.

[0014] Preferably, the discriminator of the adversarial verification module compares the difference value with a preset dynamic threshold value, and triggers feature back processing when the difference value exceeds the threshold value.

[0015] Preferably, the space-time causal graph of the causal reasoning module includes meteorological field pressure gradient nodes, wind measurement tower group spatial position nodes, and wind speed propagation path edges, and verifies the rationality of feature coupling by edge weight gradient.

[0016] Preferably, it further includes a two-stage verification loop:

[0017] The first verification stage sets a real-time evaluator after the feature fusion unit to monitor the entropy value change of the feature vector and trigger a reprocessing instruction;

[0018] The second verification stage sets a simulation predictor after the feature distillation unit to calculate the reconstruction error by inversely mapping the dimension-reduced features to the original data space.

[0019] 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.

[0020] Preferably, the meteorological channel and the wind measurement channel of the parallel space-time processing unit share a memory resource pool, and processing resource allocation is achieved through dynamic scheduling of computing tasks.

[0021] Preferably, the convolutional space-time memory network comprises a space-time convolutional layer and a gated memory unit for extracting multi-scale motion patterns in meteorological data.

[0022] Preferably, the graph pulse neural network comprises a spatial graph convolutional layer and a temporal pulse encoder for capturing dynamic association features between wind measurement tower groups.

[0023] According to an ultra-short-term wind speed prediction system combining meteorological data and multi-tower multi-layer wind measurement data, the prediction input vector is input into a wind speed prediction model to output a wind speed prediction result.

[0024] The present application provides an ultra-short-term wind speed prediction system combining meteorological data and multi-tower multi-layer wind measurement data. It has the following advantages:

[0025] The ultra-short-term wind speed prediction system combining meteorological data and multi-tower multi-layer wind measurement data effectively overcomes the real-time bottleneck of high-dimensional high-frequency meteorological and wind measurement data fusion by constructing a streaming heterogeneous data collaborative processing engine, adopting a space-time decoupled dual-channel architecture and a neuromorphic feature extraction mechanism. An adaptive data analysis matrix dynamically optimizes the data splitting path, and a parallel space-time preprocessor implements a differentiated acceleration strategy to reduce the pre-processing delay of multi-source heterogeneous data; a triple verification mechanism ensures the reliability of feature fusion and avoids the consumption of invalid computing resources. The cumulative delay problem caused by traditional serial processing procedures is solved, and the timeliness constraint of data processing links for ultra-short-term prediction is ensured.

[0026] The ultra-short-term wind speed prediction system combining meteorological data and multi-tower multi-layer wind measurement data realizes the co-evolution of data processing and prediction decision-making based on a closed-loop feedback dynamic optimization system: a two-stage verification ring monitors feature quality and output error in real time, and maintains stable system operation through a reprocessing instruction with dimension index and a backup model switching mechanism; a spatial coordinate-meteorological condition correlation graph accurately locates the prediction error source, driving the data analysis unit to optimize the vertical layering strategy and the feature extraction unit to adjust the channel sensitivity; a confidence vector fuses multi-source verification information to guide prediction period dynamic adjustment, combined with a high-frequency data acquisition mechanism to quickly respond to meteorological mutations. It breaks through the limitations of traditional systems that are "split" between "data processing" and "prediction application", and continuously improves the prediction accuracy and system fault tolerance capability in complex meteorological scenarios. BRIEF DESCRIPTION OF DRAWINGS

[0027] Figure 1A module interaction schematic diagram of a kind of ultra-short-term wind speed prediction system of the present application combined with meteorological data and multi-tower multi-layer wind data;

[0028] Figure 2 A flow schematic diagram of a kind of ultra-short-term wind speed prediction method of the present application combined with meteorological data and multi-tower multi-layer wind data. DETAILED DESCRIPTION

[0029] The technical solutions in the embodiments of the present application will be described clearly and completely below with the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, not all. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor fall within the scope of the present application.

[0030] Please refer to Figure 1 and Figure 2 The present application provides a technical solution: a kind of ultra-short-term wind speed prediction system combined with meteorological data and multi-tower multi-layer wind data, including adaptive data analysis unit, parallel space-time processing unit, neural morphological feature extraction unit, feature fusion unit and feature distillation unit connected in turn;

[0031] Adaptive data analysis unit is configured to receive meteorological grid data and wind tower data, meteorological data is split into independent sub-flow according to vertical height layer by preformat topology library, and wind data is reorganized into three-dimensional data block based on spatial topological relationship;

[0032] Parallel space-time processing unit includes meteorological channel and wind channel, meteorological channel uses interpolation algorithm driven by atmospheric motion equation to process meteorological sub-flow, and wind channel applies compression alignment technology driven by turbulent correlation model to process three-dimensional data block;

[0033] Neural morphological feature extraction unit includes convolutional space-time memory network and graph impulse neural network in parallel, which processes meteorological channel output and wind channel output respectively;

[0034] Feature fusion unit performs triple verification coupling on two feature vectors;

[0035] Feature distillation unit performs dimension reduction processing on fusion features to generate prediction input vector.

[0036] It needs to be further explained that in the specific implementation process, when the system starts, the adaptive data analysis unit receives the externally input meteorological grid data and the raw data collected by the distributed wind measurement tower group. The meteorological grid data identifies its three-dimensional structural characteristics through the pre-set format topology library, and is decoupled into independent meteorological subflows according to the vertical height layer; at the same time, the wind measurement tower group data constructs a spatial topological relationship according to the geographical coordinates, and is reorganized into a three-dimensional data block containing the tower position and height dimension. The two types of data are input into the space-time processing unit in parallel: the meteorological subflow enters the meteorological channel and is processed by an interpolation algorithm based on the atmospheric motion equation, which dynamically adjusts the interpolation weight in combination with the relationship between pressure gradient and temperature field change; the wind data block is input into the wind measurement channel, and a compression alignment technology driven by a turbulence correlation model is applied to compress the redundant information by establishing a vortex transfer function between adjacent tower layers.

[0037] The processed two-way data are respectively input into the neuromorphic feature extraction unit. The meteorological subflow is processed by a convolutional space-time memory network, and its space-time convolutional layer extracts different scales of atmospheric motion patterns, and the gate memory unit tracks the migration trajectory of the pressure system; the wind data block is processed by a graph pulse neural network, and the spatial graph convolutional layer associates the position relationship of the wind measurement tower, and the time pulse encoder captures the phase propagation characteristics of the turbulence energy. The two-way output feature vectors are input into the feature fusion unit to perform three-way verification coupling: first, the feature dimensions with a contribution degree lower than the threshold are dynamically excluded based on the information entropy change rate; then, the adversarial verification module reconstructs the feature probability distribution through the generator, and the discriminator calculates the KL divergence of the original feature and the reconstructed feature, and triggers the feature back when the divergence value exceeds the dynamically adjusted tolerance range; finally, the causal reasoning module constructs a directed graph between the pressure nodes of the meteorological field and the position nodes of the wind measurement tower, and reversely calculates the gradient loss along the wind speed propagation path, and if the loss value indicates a logical conflict, the feature reorganization is started.

[0038] The fused feature vector is compressed by the feature distillation unit: the first layer of attention mechanism focuses on the high information density dimension, and the second layer of sparse coding eliminates the linear correlation between features, generating a low-dimensional prediction input vector. In the whole processing process, the two-stage verification loop continuously monitors the data quality: when the entropy change rate of the feature fusion unit output exceeds the normal fluctuation interval, the real-time evaluator sends a reprocessing instruction to the feature extraction unit; the prediction input vector output by the feature distillation unit is mapped back to the original data space by the simulation predictor, and if the reconstruction error continues to increase, the system self-check is started.

[0039] The feature fusion unit comprises:

[0040] A feature screening module that dynamically excludes redundant feature dimensions based on the information entropy change rate;

[0041] An adversarial verification module that reconstructs the feature distribution through a generator and calculates the distribution difference value of the original feature and the reconstructed feature by a discriminator;

[0042] A causal reasoning module is configured to construct a spatiotemporal causal graph model of the meteorological data and the wind measurement data, and to verify the physical logic of feature coupling by back propagation of gradients.

[0043] It needs to 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, a triple verification coupling process is started. First, in the feature screening module, the information entropy change rate of each dimension of the meteorological feature vector is calculated, and when the entropy change rate of a specific dimension continuously falls below a set threshold, it is determined that the dimension contributes insufficiently to the dynamic representation of the wind field, and it is removed from the set of features to be fused; at the same time, the fluctuation amplitude of the entropy value of the wind measurement feature vector is monitored, and the artificial review mechanism is started for the dimension whose fluctuation range abnormally narrows.

[0044] The screened feature set is input into the adversarial verification module: the generator generates simulated wind measurement features according to the meteorological feature distribution, and the discriminator compares the KL divergence values of the real wind measurement features and the generated features; the system presets a dynamic tolerance interval, and when the KL divergence value falls below the lower limit of the interval, it is determined that the feature information is redundant, triggering a feature dimension reduction and reprocessing instruction; when the divergence value exceeds the upper limit of the interval, it is determined that the feature is mismatched, and a data reacquisition request is sent to the data analysis unit.

[0045] Finally, a directed acyclic graph structure is constructed through the causal reasoning module, which includes isobaric line nodes of the meteorological field, geographical location nodes of the wind measurement tower, and wind speed propagation path edges derived based on atmospheric dynamics; when performing back propagation calculation, if the sign correlation between the output layer wind speed prediction gradient and the input layer meteorological pressure gradient appears contradictory, it is determined that the feature coupling violates the physical law, and the conflicting feature dimension is automatically isolated and the alternative feature combination scheme is started. In the entire process, the abnormal signal output in any verification stage will interrupt the subsequent fusion operation until the self-checking of that stage is completed.

[0046] The discriminator of the adversarial verification module outputs a difference value, which is compared with a preset dynamic threshold value, and when the difference value exceeds the threshold value, feature back transmission reprocessing is triggered. It needs to be further explained that in the specific implementation process, when the adversarial verification module is running, the generator receives the meteorological feature vector as input, reconstructs the probability distribution of the wind measurement feature through a deep probability generation network, and generates a simulated wind measurement feature vector. The discriminator synchronously acquires the real wind measurement feature vector, and calculates the KL divergence values of the simulated feature and the real feature in the time-frequency spectrum feature dimension and the spatial correlation dimension, respectively. The system dynamically adjusts the tolerance interval according to the feature stability in the historical data window: when the standard deviation of the feature distribution narrows continuously for multiple periods, the interval width is compressed, and when a sudden meteorological disturbance is detected, the upper limit of the interval is expanded.

[0047] The discriminator compares the KL divergence value with the current tolerance interval, triggers a hierarchical response mechanism, including: if the divergence value is within the tolerance interval, the simulation feature is marked as valid and output to the next link; if the divergence value is below the lower limit of the interval, it is determined that the feature information is redundant, a dimension reduction instruction is sent to the feature distillation unit, and the current generator parameter is frozen; if the divergence value breaks through the upper limit of the interval, the fusion process is immediately interrupted, and a reacquisition request code is sent to the data analysis unit, which contains an abnormal feature dimension identifier to guide the data source to supplement the sampling. When the upper limit of the interval is triggered continuously, the system automatically switches to the standby generator model and resets the discriminator weight, and starts the feature channel self-diagnosis program to check the data transmission integrity. All abnormal processing records are written into the verification log for optimization of tolerance interval adjustment strategy.

[0048] The spatiotemporal causal graph of the causal reasoning module includes meteorological field pressure gradient nodes, wind tower group spatial position nodes, and wind speed propagation path edges, and the feature coupling rationality is verified by edge weight gradient. It needs to be further explained that in the specific implementation process, when the causal reasoning module is started, the pressure gradient nodes are constructed based on the current isobaric line curvature distribution, and the node weight is related to the spacing rate change between adjacent isobaric lines; at the same time, the spatial position nodes are generated according to the geographic coordinates of the wind tower group, and the node attributes include altitude and surrounding terrain shielding coefficient. The wind speed propagation path edge is created between the nodes, the direction of the edge is pointed to the position node from the pressure gradient node, and the initial weight is set according to the relationship between the pressure gradient force and the geostrophic wind speed in atmospheric dynamics.

[0049] When performing feature coupling verification, bidirectional gradient calculation is performed along the propagation path edge: forward propagation maps meteorological features to predicted wind speed distribution, and backward propagation calculates the partial derivative of the pressure gradient node to the wind speed change of the position node. When the backward gradient sign of a specific path edge conflicts with the physical law of pressure gradient, including: for example, the position node in the high pressure area presents a negative gradient but detects a positive backward weight, including: determining that the path logic is abnormal.

[0050] The system automatically isolates the feature dimensions associated with the conflict edge and activates the alternative propagation path network: the alternative network uses the path combination verified to be effective under the same working condition in history, and reconstructs the reasoning graph by attenuating the weight of the abnormal edge and strengthening the statistical significance path. The pressure gradient node that continuously triggers the conflict determination will be marked as an untrusted node and needs to be reverified in the next data processing period. The weight gradient of all path edges is stored in the causal knowledge base for optimizing the initial weight setting strategy.

[0051] It also includes a two-stage verification loop:

[0052] The first verification stage sets a real-time evaluator after the feature fusion unit to monitor the entropy value change of the feature vector and trigger a reprocessing instruction;

[0053] The second verification stage is provided behind the feature distillation unit, and the simulation predictor inversely reconstructs the reduced dimension prediction input vector into the original data space expression to calculate the grid root mean square deviation of the reconstructed meteorological data and the original meteorological data and the tower layer correlation coefficient of the reconstructed wind data and the original wind data.

[0054] It needs to be further explained that, in the specific implementation process, the two-stage verification ring is activated synchronously during system operation. The first verification stage is deployed at the output end of the feature fusion unit, and the real-time evaluator continuously monitors the information entropy value of the fusion feature vector. When the entropy value change rate is detected to exceed the historical normal fluctuation range, including, for example, the change rate increases for three sampling periods and deviates from the baseline standard deviation by more than twice, including: determining that the feature quality is abnormal, immediately sending a reprocessing instruction code to the neuromorphic feature extraction unit, which contains an abnormal entropy change dimension identifier; after receiving the instruction, the feature extraction unit restarts the feature extraction channel pointed to by the identifier, and retains the processing results of the unaffected dimensions.

[0055] The second verification stage is located behind the feature distillation unit, and the simulation predictor inversely reconstructs the reduced dimension prediction input vector into the original data space expression to calculate the grid root mean square deviation of the reconstructed meteorological data and the original meteorological data and the tower layer correlation coefficient of the reconstructed wind data and the original wind data. If the meteorological deviation continues to increase and the wind correlation coefficient continues to decrease, start the system self-checking protocol: first isolate the feature distillation unit output port, enable the standby dimension reduction model to generate a temporary prediction input vector; at the same time, trigger the data analysis unit to perform timestamp integrity verification on the original data to exclude time domain misalignment caused by data transmission delay.

[0056] The prediction results generated during the self-checking period are marked with a confidence identifier, and the isolation state is released after the deviation and correlation coefficient return to the normal threshold. The two-stage verification log is synchronized in real time to the system console to form a closed-loop diagnosis report.

[0057] The real-time evaluator is configured to send a reprocessing instruction to the neuromorphic feature extraction unit when the feature vector entropy value change rate exceeds the set range. It needs to be further explained that, in the specific implementation process, the real-time evaluator continuously tracks the feature vector entropy value change rate output by the feature fusion unit during operation, and maintains the historical normal fluctuation baseline through the dynamic learning module. When the entropy change rate of a specific dimension continuously exceeds the baseline range, a multi-level judgment process is started: first, verify whether the abnormal dimension is associated with a recent meteorological disturbance event in time and space, if there is an association, then relax the judgment threshold by one level; if it is an isolated anomaly, further analyze the information contribution weight of the dimension in the feature vector, and only mark the dimension with a weight below the threshold without triggering the instruction.

[0058] When necessary, generate reprocessing instruction code containing the abnormal dimension index, and send it to the neuromorphic feature extraction unit during the non-peak period. After receiving the instruction, the feature extraction unit performs a step-by-step response: only restart the feature extraction channel pointed to by the index, and keep the processing results of the remaining channels; the restart process uses a gradual initialization strategy, first loads the historical feature mean of the first three periods before the abnormal dimension as the initial state, and then gradually fuses real-time data streams.

[0059] If the same dimension triggers reprocessing continuously, the system automatically raises the priority of the channel in the computing resource pool and generates a dimension health report to push to the operation and maintenance interface. All reprocessing operations record the dimension index and response delay data for optimizing the instruction sending strategy.

[0060] The meteorological channel and the wind measurement channel of the parallel space-time processing unit share the memory resource pool, and the processing resource allocation is realized through dynamic scheduling of computing tasks. It needs to be further explained that during the operation of the parallel space-time processing unit, the meteorological channel and the wind measurement channel realize the cooperative allocation of computing resources through the shared memory resource pool. When the resource pool is initialized, the basic computing units are reserved for the minimum guarantee of the double channels, and the remaining resources form a dynamic allocation pool. The resource scheduler continuously monitors the processing state of the double channels: when the meteorological channel detects sudden pressure gradient changes, it sends a resource preemption request code to the scheduler, which contains the number of computing units required and the expected duration; after the scheduler verifies the reasonableness of the request, it temporarily allocates resources from the dynamic pool to the meteorological channel, and starts the buffer queue mechanism for the wind measurement channel to temporarily store the data blocks to be processed.

[0061] If the wind measurement channel triggers a load alarm due to an increase in tower layer data, the scheduler uses a gradual resource release strategy, including: preferentially reducing the resource allocation of non-core interpolation operations of the meteorological channel, and transferring computing units to the wind measurement channel in batches.

[0062] 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 the historical model parameters of the same working condition to preheat the newly allocated resources. When the load of the double channels returns to the balanced range, the scheduler starts the resource recovery protocol: gradually recovers the temporarily allocated units and releases the wind measurement buffer queue, and verifies the timeliness of the data blocks during the processing period, and starts the acceleration processing thread for the data blocks that are close to the timeout. All resource scheduling records generate optimization logs, which are used to train the load prediction model to improve the forward-looking of scheduling.

[0063] The convolutional spatio-temporal memory network comprises a spatio-temporal convolutional layer and a gated memory unit, which are used to extract multi-scale motion patterns in meteorological data. It should be further explained that, in the implementation process, when the convolutional spatio-temporal memory network processes meteorological subflows, the spatio-temporal convolutional layer adopts a multi-scale sliding window sampling mechanism. When inputting the pressure field data, first, the regional pressure gradient change direction is detected, and a long strip-shaped convolutional kernel is deployed along the gradient direction to capture the migration characteristics of the pressure system; for temperature field data, a ring-shaped convolutional kernel is used to identify the evolution pattern of the thermal center.

[0064] The gated memory unit dynamically tracks the evolution trajectory of key meteorological elements: in the initialization stage, the region with continuously enhanced pressure gradient is marked as the focus region, and the memory unit is allocated with additional storage weight; when it is detected that the pressure gradient change rate of the focus region exceeds the historical benchmark, the convolutional 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.

[0065] The memory unit regularly evaluates the storage state, and attenuates the memory weight of the meteorological pattern that has not been activated for three consecutive periods, so as to release space for newly formed disturbance systems. When processing sudden meteorological events, the network automatically switches to a backup memory channel: the channel pre-stores typical weather process models, such as frontal passage, and quickly loads the corresponding parameters through similarity matching. The weight update records of all convolutional kernels are input into the feature analysis module, which is used to optimize the initial convolutional kernel configuration strategy.

[0066] The graph pulse neural network comprises a spatial graph convolutional layer and a temporal pulse encoder, which are used to capture dynamic correlation features between wind measurement tower groups. It should be further explained that, in the implementation process, when the graph pulse neural network is initialized, a spatial graph structure is constructed according to the three-dimensional coordinates of the wind measurement tower group, and the connection weight between nodes is calibrated according to the terrain shielding effect and the historical turbulence correlation intensity. When processing wind data blocks, the spatial graph convolutional layer transmits turbulence energy features along the connection edges, and increases the flow compensation coefficient for the tower nodes in the valley terrain area.

[0067] The temporal pulse encoder adopts a phase modulation mechanism: when it is detected that the dominant wind direction is stable and continuous, the feature propagation phase is locked to reduce the calculation overhead; when the wind direction mutates by more than a threshold angle, a multi-phase parallel encoding is started to capture the turbulence transient characteristics. The network dynamically monitors the energy transmission efficiency between nodes, and performs weight reduction processing on the connection edges with continuously increasing transmission delay, and activates the backup topology path, including: the path is based on the connection mode verified to be effective under the historical similar wind direction working condition.

[0068] When processing strong shear layer data, a vertical momentum correction factor is automatically injected to suppress feature distortion. The pulse activation mode of all nodes is recorded into a feature evolution graph, which is used to optimize the initial connection weight allocation strategy.

[0069] According to an ultra-short-term wind speed prediction system combining meteorological data and multi-tower multi-layer wind measurement data, a prediction input vector is input into a wind speed prediction model to output a wind speed prediction result. It needs to be further explained that in the specific implementation process, when the prediction input vector is input into the wind speed prediction model, the model preferentially loads the parameter set matching the current atmospheric stability level. The processing process continuously monitors the activity index of each dimension in the input vector, and when it is detected that the terrain disturbance related dimension continuously dominates the feature space, the decision weight of the local terrain correction sub-module is automatically enhanced; if the thermal convection dimension activity suddenly increases, the boundary layer parameterization scheme is activated to suppress excessive response.

[0070] The model outputs a prediction result and generates a confidence evaluation vector at the same time, which integrates three pieces of information: the triple verification state code of the feature fusion unit, the abnormal marker of the double-stage verification ring, and the residual analysis result of the prediction model itself. When the confidence evaluation shows that the meteorological system is in a rapid evolution stage, the system automatically shortens the time window length of the next prediction period and starts a high-frequency data acquisition mode to supplement the observation blind area.

[0071] After all the prediction results are compared with the measured data, an error distribution graph is generated, which marks the spatial coordinates and corresponding meteorological conditions of the area with continuous high error. This information is fed back to the adaptive data analysis unit in real time to optimize the vertical layering strategy, and at the same time guides the neuromorphic feature extraction unit to adjust the sensitivity parameters of the feature extraction channel. The intermediate data of the entire prediction process is stored in a rolling buffer area to support the reconstruction of the feature processing state at any time point during backtracking analysis.

[0072] It needs to be further explained that in the specific implementation process, when the system is started, the adaptive data analysis unit receives meteorological grid data and wind measurement tower group raw data. The meteorological data is identified by a pre-set format analysis library for its three-dimensional structural features, and is split into independent sub-streams according to vertical height layers, and the splitting process dynamically adjusts the layer density with reference to the atmospheric boundary layer height. The wind measurement tower group data constructs a spatial topology network according to geographical coordinates, and reorganizes the discrete tower layer observation values into three-dimensional data blocks with spatial position identifiers. When the two types of data are input into the space-time processing unit in parallel: the meteorological sub-stream enters the meteorological processing channel, and an interpolation algorithm based on the atmospheric motion equation is used, which combines the pressure gradient and temperature advection relationship to dynamically correct the interpolation weight; the wind data block enters the wind measurement processing channel, and a compression alignment technology driven by a turbulence correlation model is applied to eliminate data redundancy by establishing a vortex transfer function between adjacent tower layers.

[0073] The processed two-way data enters the neuromorphic feature extraction link. The meteorological sub-flow is processed by a convolutional spatio-temporal memory network, whose multi-scale convolution kernel captures the moving characteristics of weather systems along the pressure gradient direction, and the gating memory unit tracks the evolution trajectory of the meteorological elements in the high variability area. The wind data block is processed by a graph pulse neural network, and the spatial graph convolution layer fuses the terrain shielding effect to calculate the node correlation degree. The time pulse encoder switches the encoding phase according to the wind direction stability. The feature vectors output by the two paths are input into the feature fusion unit to perform triple verification: first, based on the information entropy change rate, dynamically screen out feature dimensions with insufficient contribution, and the screening process refers to the feature effectiveness record under the same working condition; 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 interval, the feature back mechanism is triggered; finally, the causal reasoning module constructs a directed graph model between the meteorological pressure nodes and the wind measurement location nodes, and verifies the gradient sign consistency along the wind speed propagation path, and detects physical contradictions such as negative feedback of wind speed in high pressure area when detecting conflict features.

[0074] The fused features are compressed into low-dimensional vectors through a distributed distillation pipeline. This process uses attention mechanism to focus on high information dimension, and eliminates linear correlation between features through sparse coding. The distilled prediction input vector is input into the wind speed prediction model, which loads the corresponding parameter set according to the current atmospheric stability, and outputs the prediction result and the confidence vector simultaneously. The confidence integrates three pieces of information: feature fusion verification state identifier, data processing anomaly marker, and model residual analysis conclusion. When the confidence shows that the meteorological system is evolving rapidly, the system automatically shortens the prediction period and starts high-frequency data acquisition.

[0075] The entire process is monitored by a two-stage verification loop: in the first stage, a real-time evaluator is set after feature fusion. When the feature vector entropy change rate 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 prediction input vector is output, it is mapped back to the original data space through a simulation predictor, and the correlation coefficient between the meteorological grid reconstruction bias and the wind data is calculated. When the anomaly persists, a backup dimension reduction model is enabled and the data transmission integrity is verified. All intermediate data in the processing are stored in a rolling buffer area 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 to guide the data analysis unit to optimize the vertical layering strategy and the feature extraction unit to adjust the channel sensitivity.

[0076] An ultra-short-term wind speed prediction method combining meteorological data and multi-tower multi-layer wind data, comprising the following steps:

[0077] Step S1: The adaptive data analysis unit receives meteorological grid data and raw data of the wind tower group, and splits the meteorological data into independent sub-flows according to the vertical height layer, and reorganizes the wind data into three-dimensional blocks based on spatial topology;

[0078] Step S2: Parallel spatio-temporal processing unit starts dual-channel processing: meteorological channel adopts dynamic interpolation algorithm driven by atmospheric motion equation, and wind measurement channel applies compression alignment technology driven by turbulent correlation model;

[0079] Step S3: Neuromorphic feature extraction unit synchronously processes dual-channel data: convolutional spatio-temporal memory network extracts meteorological multi-scale motion features, and graph impulse neural network captures dynamic correlation features of wind tower group;

[0080] Step S4: Feature fusion unit performs triple verification coupling: low-contribution feature dimensions are filtered based on information entropy change rate, an adversarial verification module reconstructs feature distribution through a generator, a discriminator compares distribution difference values and triggers dynamic tolerance response, and a causal reasoning module constructs a directed graph of meteorological pressure nodes and wind measurement location nodes to verify the physical consistency of gradient signs;

[0081] Step S5: Distributed feature distillation pipeline compresses and fuses features: attention mechanism is used to focus on high-information-dimension, and sparse coding is used to eliminate linear correlation between features;

[0082] Step S6: Wind speed prediction model loads atmospheric stability matching parameter set, and outputs prediction results to generate confidence vector with triple verification state;

[0083] Step S7: Two-stage verification loop monitors in real time: real-time evaluator monitors entropy change rate after feature fusion, and triggers reprocessing instruction with dimension index when abnormal; after the output of the prediction input vector, the simulation predictor calculates the reconstruction bias and correlation coefficient in reverse, and starts the standby dimension reduction model when the abnormality persists;

[0084] Step S8: Prediction error analysis generates a space coordinate-meteorological condition correlation graph, which is fed back to the data analysis unit to optimize the vertical stratification strategy and adjust the channel sensitivity of the feature extraction unit;

[0085] Step S9: Rolling buffer stores intermediate data throughout the process, supporting any node processing state rollback;

[0086] Step S10: Dynamically adjust the prediction period according to the confidence vector: shorten the period and activate high-frequency data acquisition when the meteorological system evolves rapidly.

[0087] By constructing a streaming heterogeneous data collaborative processing engine, using a spatio-temporal decoupling dual-channel architecture and neuromorphic feature extraction mechanism, the real-time bottleneck of high-dimensional and high-frequency meteorological and wind data fusion is effectively overcome. The adaptive data analysis matrix dynamically optimizes the data splitting path, and the parallel spatio-temporal preprocessor implements differentiated acceleration strategies to reduce the preprocessing delay of multi-source heterogeneous data; the triple verification mechanism ensures the reliability of feature fusion and avoids the consumption of invalid computing resources. The cumulative delay problem caused by traditional serial processing flow is solved, ensuring that the data processing link meets the timeliness constraints of ultra-short-term prediction.

[0088] The dynamic optimization system based on closed-loop feedback realizes the synergistic evolution of data processing and predictive decision-making: the two-stage verification loop monitors the feature quality and output error in real time, and maintains stable operation of the system through reprocessing instructions with dimension index and backup model switching mechanism; the spatial coordinate-weather condition correlation diagram accurately locates the prediction error source, drives the data analysis unit to optimize the vertical stratification strategy and the feature extraction unit to adjust the channel sensitivity; the confidence vector fuses multi-source verification information to guide the dynamic adjustment of the prediction period, combined with the high-frequency data acquisition mechanism to quickly respond to meteorological mutations. Breakthrough the limitations of traditional system "data processing-prediction application" fragmentation, continuously improve the prediction accuracy and system fault tolerance ability under complex weather conditions.

[0089] It should be noted that, in this document, the terms such as first and second are used merely to distinguish one entity or action from another, and do not necessarily require or imply that these entities or actions exist in any actual relationship or order. Moreover, the terms "include", "contain" or any other variant thereof are intended to cover non-exclusive inclusion, so that the process, method, article or device including a series of elements not only includes those elements, but also includes other elements not explicitly listed or inherent to such process, method, article or device. Without more limitations, the element defined by the statement "including a" does not exclude the presence of other identical elements in the process, method, article or device including the element.

[0090] Although embodiments of the present application have been shown and described, it will be understood by those having ordinary skill in the art that various changes, modifications, substitutions and alterations can be made thereto without departing from the principles and spirit of the present application, the scope of which is defined by the appended claims and their equivalents.

Claims

1. An ultra-short term wind speed prediction system that combines meteorological data with multi-tower multi-layer wind measurement data, characterized by, The adaptive data analysis unit, the parallel space-time processing unit, the neuromorphic feature extraction unit, the feature fusion unit and the feature distillation unit are sequentially connected. The adaptive data analysis unit is configured to receive meteorological grid data and wind tower group data, and split the meteorological data into independent sub-flows according to vertical height layers through a preset format topology library, and reorganize the wind data into three-dimensional data blocks based on spatial topological relationship. The parallel space-time processing unit includes a meteorological channel and a wind channel, the meteorological channel uses an interpolation algorithm driven by an atmospheric motion equation to process the meteorological sub-flow, and the wind channel uses a compression alignment technology driven by a turbulence correlation model to process the three-dimensional data blocks. The neuromorphic feature extraction unit includes a convolutional space-time memory network and a graph impulse neural network connected in parallel, which respectively process the output of the meteorological channel and the output of the wind channel. The feature fusion unit performs triple verification coupling on two feature vectors, and includes: A feature screening module that dynamically screens out redundant feature dimensions based on information entropy change rate; An adversarial verification module that reconstructs feature distribution by a generator and calculates the distribution difference value of the original feature and the reconstructed feature by a discriminator; A causal reasoning module that constructs a space-time causal graph model of meteorological data and wind data, and verifies the physical logic of feature coupling through back propagation gradient; The feature distillation unit performs dimension reduction processing on the fused features to generate a prediction input vector; The prediction input vector is input into a wind speed prediction model to output a wind speed prediction result.

2. The system for ultra-short term wind speed prediction combining meteorological data with multi-tower multi-layer wind data according to claim 1, characterized in that: The discriminator of the adversarial verification module compares the difference value with a preset dynamic threshold value, and triggers feature back processing when the difference value exceeds the threshold value.

3. The system of claim 1, wherein the system is characterized by: The space-time causal graph of the causal reasoning module includes meteorological field pressure gradient nodes, wind tower group spatial position nodes and wind speed propagation path edges, and verifies the rationality of feature coupling through edge weight gradient.

4. The system of claim 1, wherein the system is characterized by: It also includes a two-stage verification loop: A real-time evaluator is arranged after the feature fusion unit in the first verification stage to monitor the entropy value change of the feature vector and trigger a reprocessing instruction; An analog predictor is arranged after the feature distillation unit in the second verification stage to calculate the reconstruction error by mapping the reduced features back to the original data space.

5. The system for ultra-short term wind speed prediction combining meteorological data with multi-tower multi-layer wind data according to claim 4, characterized in that: The real-time evaluator is configured to send a reprocessing instruction to the neuromorphic feature extraction unit when the feature vector entropy value change rate exceeds a set range.

6. The system for ultra-short term wind speed prediction combining meteorological data with multi-tower multi-layer wind data according to claim 1, characterized in that: The meteorological channel and the wind channel of the parallel space-time processing unit share a memory resource pool, and the processing resource allocation is realized through dynamic scheduling of computing tasks.

7. The system for ultra-short term wind speed prediction combining meteorological data with multi-tower multi-layer wind data according to claim 1, characterized in that: The convolutional space-time memory network includes a space-time convolution layer and a gated memory unit, which are used to extract multi-scale motion patterns in meteorological data.

8. The system for ultra-short term wind speed prediction combining meteorological data with multi-tower multi-layer wind data according to claim 1, wherein: The graph impulse neural network includes a spatial graph convolution layer and a temporal pulse encoder, which are used to capture dynamic correlation features between wind tower groups.

Images