Long short-term memory random forest wind speed prediction method and system based on terrain influence

Abstract

The application discloses a long short-term memory random forest wind speed prediction method and system based on terrain influence, relates to the technical field of data processing, and comprises the following steps: obtaining historical wind speed, environment and terrain characteristics of each wind measurement node and normalizing; inputting time sequence characteristics into a long short-term memory network to obtain a predicted value and a short-term fluctuation characteristic component; inputting the component and the environment and terrain characteristics into a random forest model to obtain a predicted value; obtaining a dynamic weight according to a deviation between the two predicted values and an adaptive deviation threshold value, and generating a fused predicted value; obtaining an aerodynamic equivalent distance based on vertical height difference and horizontal distance, and performing spatial interpolation to obtain a target node wind speed forecast. The application solves the problems that high-frequency turbulence cannot be decoupled, weight distribution is unreasonable and spatial interpolation ignores physical barriers under complex terrain, and improves wind speed prediction accuracy.

Description

Technical Field

[0001] This invention relates to the field of data processing technology, specifically to a method and system for predicting wind speed in long short-term memory random forests based on terrain influence. Background Technology

[0002] In power grid disaster prevention and mitigation and wind power dispatch assessment, obtaining multi-scale spatiotemporal distribution characteristics of wind speed along power lines is a prerequisite for ensuring stable system operation. However, in complex geographical areas with undulating terrain and large elevation differences, the wind field structure is affected by airflow dynamics and thermodynamic distribution, resulting in strong spatial non-uniformity and temporal non-stationarity of local wind speed variations. Existing prediction technologies have certain limitations in dealing with such microclimate environments.

[0003] First, existing time-series deep learning models tend to fit the low-frequency stable trend of macro-meteorology when extracting wind speed evolution patterns. When faced with high-frequency mechanical turbulence induced by complex micro-topography, they struggle to effectively decouple and isolate the short-term fluctuation features that play a major role in wind speed abrupt changes, leading to response lag and feature smoothing phenomena in extreme gust predictions. Second, in the nonlinear mapping of multi-dimensional environmental features and wind speed, and in the process of multi-model fusion, existing solutions often use static weight allocation or rely on the absolute difference of model outputs to determine prediction reliability. When real-time airflow impacts a steep windward slope, it triggers wind speed fluctuations. In this case, data discrepancies between different prediction models are normal responses to mechanical turbulence, but existing evaluation mechanisms may classify this as model failure, inducing tolerance penalties and weight decay, causing the fusion result to deviate from the actual wind field state. Finally, when interpolating data from discrete wind measurement nodes to target spatial locations without monitoring equipment, existing spatial reconstruction algorithms rely on two-dimensional horizontal straight-line distance as the interpolation attenuation scale, failing to consider the physical blocking effects of mountain barriers and deep valleys on airflow in three-dimensional space. This approach may result in two nodes located on opposite sides of a ridge or with a significant elevation difference and close horizontal distance but isolated by airflow being assigned high spatial correlation weights, leading to distortions in cross-terrain wind speed projections. Therefore, there is an urgent need for a prediction scheme that can combine three-dimensional spatial physical barriers with airflow turbulence mechanisms to overcome the challenges of wind speed prediction under complex micro-topographic conditions. Summary of the Invention

[0004] In response to the technical problems described in the background art, the present invention provides a long short-term memory random forest wind speed prediction method and system based on terrain influence. It solves the technical problems of wind speed spatiotemporal prediction distortion and insufficient accuracy caused by existing wind speed prediction models in complex micro-topography (such as mountains and canyons) and local microclimate environments, due to their inability to effectively decouple high-frequency mechanical turbulence characteristics, unreasonable fusion weight allocation, and spatial interpolation ignoring three-dimensional terrain physical barriers.

[0005] A method for wind speed prediction based on long short-term memory (LSTM) random forest based on terrain influence is proposed. The method includes: acquiring historical wind speed sequences, environmental feature parameters, and terrain feature parameters for each wind measurement node, and performing normalization processing; constructing a temporal feature vector based on the historical wind speed sequence, inputting the temporal feature vector into an LTM network to obtain LTM network prediction values, and extracting short-term fluctuation feature components within the LTM network; concatenating the short-term fluctuation feature components, environmental feature parameters, and terrain feature parameters to obtain a model input vector, and inputting the model input vector into a random forest model to obtain random forest prediction values; obtaining the model bias based on the LTM network prediction values ​​and the random forest prediction values, and obtaining an adaptive bias threshold based on the difference between adjacent data in the historical wind speed sequence and the terrain feature parameters; obtaining dynamic weight coefficients based on the adaptive bias threshold and the model bias, and generating fused prediction values ​​for the wind measurement nodes based on the dynamic weight coefficients; obtaining the aerodynamic equivalent distance based on the vertical height difference and horizontal distance between the target location node and the wind measurement node, and performing spatial interpolation processing on the fused prediction values ​​based on the aerodynamic equivalent distance to obtain the wind speed forecast results for the target location node.

[0006] Optionally, the short-term fluctuation feature components within the Long Short-Term Memory Network are extracted, including: extracting the forget gate output vector and the candidate unit state vector at the last time step of the Long Short-Term Memory Network; obtaining the element-wise product of the forget gate output vector and the candidate unit state vector, and determining the element-wise product as the short-term fluctuation feature components.

[0007] Optionally, inputting the model input vector into the random forest model to obtain the random forest prediction value includes: calculating the model input vector using multiple decision trees in the random forest model to obtain the output value of each decision tree; obtaining the arithmetic mean of all output values ​​of the multiple decision trees, and determining the arithmetic mean as the random forest prediction value.

[0008] Optionally, an adaptive deviation threshold is obtained based on the difference between adjacent data in the historical wind speed sequence and terrain feature parameters, including: extracting the maximum absolute value of the difference between adjacent data in the historical wind speed sequence and using the maximum value as the tolerance threshold base; obtaining the measured wind direction angle of the wind measurement node, as well as the normal azimuth angle and tilt angle in the terrain feature parameters; when the cosine value of the difference between the measured wind direction angle and the normal azimuth angle is positive, multiplying the cosine value by the sine value of the tilt angle to obtain the penalty multiple; when the cosine value is not positive, the penalty multiple is set to zero; adding the value one to the penalty multiple to obtain the scaling factor, and multiplying the tolerance threshold base by the scaling factor to obtain the adaptive deviation threshold.

[0009] Optionally, dynamic weight coefficients are obtained based on the adaptive bias threshold and the model bias, and fused prediction values ​​of wind measurement nodes are generated based on the dynamic weight coefficients, including: obtaining the maximum value between the adaptive bias threshold and the model bias; adding the maximum value to the adaptive bias threshold to obtain a dynamic denominator; dividing the adaptive bias threshold by the dynamic denominator to obtain the first dynamic weight coefficient of the long short-term memory network; subtracting the first dynamic weight coefficient from the first value to obtain the second dynamic weight coefficient of the random forest model; multiplying the prediction value of the long short-term memory network with the first dynamic weight coefficient to obtain a first product; multiplying the prediction value of the random forest with the second dynamic weight coefficient to obtain a second product; and adding the first product and the second product to obtain the fused prediction value.

[0010] Optionally, based on the vertical elevation difference and horizontal distance between the target location node and the wind measurement node, an aerodynamic equivalent distance is obtained. Spatial interpolation is then performed on the fused prediction value based on this aerodynamic equivalent distance to obtain the wind speed forecast result for the target location node. This includes: obtaining the two-dimensional horizontal straight-line distance between the target location node and the wind measurement node, and the absolute vertical elevation difference between the elevation of the target location node and the elevation of the wind measurement node; adding the two-dimensional horizontal straight-line distance and the absolute vertical elevation difference to obtain the aerodynamic equivalent distance; multiplying the fused prediction value of each wind measurement node by the negative square of the corresponding aerodynamic equivalent distance and summing the results to obtain the interpolation numerator; summing the negative squares of the aerodynamic equivalent distances of each wind measurement node to obtain the interpolation denominator; and dividing the interpolation numerator by the interpolation denominator to obtain the wind speed forecast result for the target location node.

[0011] Optionally, historical wind speed sequences, environmental characteristic parameters, and terrain characteristic parameters of each wind measurement node are obtained and normalized, including: obtaining initial wind speed data and initial environmental data of each wind measurement node; using linear interpolation to fill in missing time node data in the initial wind speed data and initial environmental data to obtain a continuous data sequence; subtracting the minimum value in the continuous data sequence from the currently collected value in the continuous data sequence, and dividing by the difference between the maximum and minimum values ​​in the continuous data sequence to obtain the normalized values ​​corresponding to the historical wind speed sequence and environmental characteristic parameters.

[0012] Optionally, after obtaining the wind speed forecast results for the target location nodes, the method further includes: when there are multiple target location nodes distributed across the target spatial grid, extracting the maximum and average wind speed values ​​from the wind speed forecast results of all target location nodes at the current prediction time; continuously extracting the maximum and average wind speed values ​​at multiple prediction times to generate a wind speed forecast time series curve; comparing the wind speed forecast time series curve with the pre-stored safe extreme value range, and generating wind speed warning information containing the spatial coordinate range and the expected occurrence time when the maximum wind speed value is outside the safe extreme value range.

[0013] A wind speed prediction system based on terrain influence and a long short-term memory random forest is also provided, including: a parameter acquisition module for acquiring historical wind speed sequences, environmental feature parameters, and terrain feature parameters for each wind measurement node, and performing normalization processing; a feature decoupling module for constructing a time-series feature vector based on the historical wind speed sequence, inputting the time-series feature vector into the long short-term memory network, obtaining the predicted value of the long short-term memory network, and extracting the short-term fluctuation feature components within the long short-term memory network; and a model mapping module for concatenating the short-term fluctuation feature components, environmental feature parameters, and terrain feature parameters to obtain a model input vector, and inputting the model input vector into the random forest. The model acquires random forest predictions; the error allocation module is used to obtain the model bias based on the predictions of the long short-term memory network and the random forest, obtain an adaptive bias threshold based on the difference between adjacent data in the historical wind speed sequence and terrain feature parameters, obtain dynamic weight coefficients based on the adaptive bias threshold and the model bias, and generate fused predictions for the wind measurement nodes based on the dynamic weight coefficients; the spatial reconstruction module is used to obtain the aerodynamic equivalent distance based on the vertical height difference and horizontal distance between the target location node and the wind measurement node, perform spatial interpolation processing on the fused predictions based on the aerodynamic equivalent distance, and obtain the wind speed forecast results for the target location node.

[0014] Optionally, when the error allocation module obtains the adaptive deviation threshold based on the difference between adjacent data in the historical wind speed sequence and the terrain feature parameters, it specifically performs the following steps: extracting the maximum absolute value of the difference between adjacent data in the historical wind speed sequence and using the maximum value as the tolerance threshold base; obtaining the measured wind direction angle of the wind measurement node, as well as the normal azimuth angle and tilt angle in the terrain feature parameters; when the cosine value of the difference between the measured wind direction angle and the normal azimuth angle is positive, multiplying the cosine value by the sine value of the tilt angle to obtain the penalty multiple; when the cosine value is not positive, setting the penalty multiple to zero; adding the value one to the penalty multiple to obtain the scaling factor, and multiplying the tolerance threshold base by the scaling factor to obtain the adaptive deviation threshold.

[0015] The beneficial effects of this invention are reflected in: In the entire topographic-influence-based long short-term memory random forest wind speed prediction method, firstly, by decoupling and extracting short-term fluctuation feature components within the long short-term memory network, the long-term stable meteorological background is removed. This allows the feature vectors input into the subsequent random forest model to focus on high-frequency wind speed disturbances induced by micro-topography, overcoming the smoothing failure problem of deep networks in microclimate gust prediction and enhancing the sensitivity to capturing local airflow abrupt changes. Secondly, a dynamic error allocation mechanism driven by the physical form of the data is established, replacing fixed preset parameters. By extracting the maximum difference between adjacent data in the historical wind speed sequence as the fluctuation baseline, and introducing the cosine of the angle between the wind direction and the normal of the topographic slope as the trigger condition for topographic penalty, the tolerance threshold is adaptively amplified when airflow impacts the windward slope and induces mechanical turbulence. This dynamic computation mechanism considers the model prediction differences caused by local airflow fluctuations, reduces the weight decay caused by excessive judgment errors, and improves the stability of multi-model fusion results. Finally, in the spatial wind field reconstruction stage, the absolute vertical height difference between the target location node and the wind measurement node is converted into horizontal resistance, and an equivalent penalty distance is constructed. This process converts the spatial vertical height difference and horizontal distance into an equivalent distance. When two points are close in horizontal distance but have a height difference, the equivalent distance is increased and the interpolation weight is reduced, which reduces the interpolation error caused by crossing the ridge. This makes the wind speed forecast results extrapolated to the entire line fit the attenuation characteristics of the actual airflow, and provides reliable data support for power grid disaster prevention and early warning. Attached Figure Description

[0016] To more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the accompanying drawings used in the description of the specific embodiments or the prior art will be briefly introduced below. In all the drawings, similar elements or parts are generally identified by similar reference numerals. In the drawings, the elements or parts are not necessarily drawn to scale.

[0017] Figure 1 This is a schematic diagram illustrating the steps of the long short-term memory random forest wind speed prediction method based on terrain influence of the present invention. Figure 2 This is a schematic diagram of a portion of step S1 in the Long Short-Term Memory Random Forest Wind Speed ​​Prediction Method Based on Topographic Influence of the present invention. Figure 3 This is a schematic diagram of part of step S2 in the Long Short-Term Memory Random Forest Wind Speed ​​Prediction Method Based on Topographic Influence of the present invention. Figure 4 This is a schematic diagram of a portion of step S3 in the Long Short-Term Memory Random Forest Wind Speed ​​Prediction Method Based on Topographic Influence of the present invention. Figure 5 This is a schematic diagram of part of step S4 in the Long Short-Term Memory Random Forest Wind Speed ​​Prediction Method Based on Topographic Influence of the present invention. Detailed Implementation

[0018] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. It should be understood that the described embodiments are only some embodiments of the present invention, and not all embodiments. The components of the embodiments of the present invention described and shown in the accompanying drawings can generally be arranged and designed in various different configurations.

[0019] Therefore, the following detailed description of the embodiments of the invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but merely to illustrate selected embodiments of the invention. All other embodiments obtained by those skilled in the art based on the embodiments of the invention without inventive effort are within the scope of protection of the invention.

[0020] It should be noted that similar labels and letters in the following figures indicate similar items; therefore, once an item is defined in one figure, it does not need to be further defined and explained in subsequent figures. Furthermore, the terms "first," "second," etc., are used only to distinguish descriptions and should not be construed as indicating or implying relative importance.

[0021] This invention provides a method for predicting wind speed in long short-term memory stochastic forests based on topographic influence, such as... Figure 1 As shown, in one specific embodiment, the method includes: S1. Obtain the historical wind speed sequence, environmental characteristic parameters, and terrain characteristic parameters of each wind measurement node, and perform normalization processing; S2. Construct a time-series feature vector based on historical wind speed sequences. Input the time-series feature vector into a Long Short-Term Memory (LSTM) network to obtain the LTM network's predicted values, and extract the short-term fluctuation feature components within the LTM network. Here, the short-term fluctuation feature components refer to the tensor information decoupled from the deep state of the LTM network that only reflects the recent wind speed change trend. Concatenate the short-term fluctuation feature components, environmental feature parameters, and terrain feature parameters to obtain the model input vector. Input the model input vector into a random forest model to obtain the random forest's predicted values. S3. Obtain the model bias based on the predictions from the Long Short-Term Memory network and the Random Forest, and obtain an adaptive bias threshold based on the differences between adjacent data in the historical wind speed sequence and terrain feature parameters. Here, the adaptive bias threshold refers to the model's fault tolerance boundary dynamically calculated based on the coupled physical effects of local terrain and real-time wind direction; obtain dynamic weight coefficients based on the adaptive bias threshold and the model bias, and generate fused prediction values ​​for the wind measurement nodes based on the dynamic weight coefficients; S4. Based on the vertical elevation difference and horizontal distance between the target location node and the wind measurement node, obtain the aerodynamic equivalent distance. Here, the aerodynamic equivalent distance refers to the comprehensive impedance path obtained by converting the vertical elevation difference between two points into horizontal air resistance in space. Based on the aerodynamic equivalent distance, perform spatial interpolation processing on the fused prediction value to obtain the wind speed forecast result of the target location node.

[0022] In this embodiment, it should be noted that in S1, multi-source features are first acquired and normalized. Specifically, historical wind speed sequences, environmental feature parameters, and terrain feature parameters of each wind measurement node are acquired. After acquiring the initial data, linear interpolation is used to fill in the missing time node data in the initial wind speed and initial environmental data to ensure the continuity of the data sequence in the time dimension. Subsequently, normalization is performed by subtracting the minimum value of the sequence from the currently collected value in the continuous data sequence and dividing by the difference between the maximum and minimum values, thereby mapping all physical quantities to a unified dimensionless interval. For example, if the historical wind speed interval of a certain wind measurement node is 0 to 35.0 m / s, and the current measured wind speed is 15.2 m / s, the normalized value will be calculated as 0.4343. This processing eliminates the dimensional differences of various meteorological parameters, making them meet the input requirements of the model and facilitating the stable convergence and training of the model.

[0023] In S2, a Long Short-Term Memory (LSTM) network is used for prediction, and the underlying short-term fluctuation features are extracted and decoupled. A time-series feature vector is constructed based on the normalized historical wind speed sequence and input into the LSM network for forward propagation to obtain the network's predicted wind speed value for the current node. For example, for the measured wind speed node of 15.2 m / s, the network output prediction is assumed to be 16.5 m / s. Subsequently, in the last time step within the network, the forget gate output vector and the candidate unit state vector are extracted, and their element-wise product is calculated. Assuming that on a certain feature dimension, the forget gate output value is 0.85 and the candidate unit state value is 0.60, the product yields a value of 0.51, which is identified as the short-term fluctuation feature component. Finally, this component is concatenated with environmental and terrain features to obtain the model input vector, which is then input into a random forest model to obtain a predicted value, assumed to be 13.5 m / s. This processing filters out the stable macro-climate background through feature product and extracts high-frequency wind speed disturbance features induced by micro-topography, enhancing the model's ability to capture sudden changes in local airflow.

[0024] In S3, a fault tolerance mechanism dominated by terrain flow field effects is constructed, and error weights are dynamically allocated. First, the absolute difference between the predictions of the Long Short-Term Memory (LSTM) network and the Random Forest model is calculated as the model bias; the difference between the two predictions is 3.0 m / s. The maximum absolute value of the difference between the past adjacent data of a node is extracted as the tolerance threshold base, assumed to be 2.8 m / s. Combining the measured wind direction angle of 270°, normal azimuth angle of 280°, and tilt angle of 30°, the cosine of the angle between the wind direction and the normal is calculated to be approximately 0.9848, which is greater than 0. This cosine is multiplied by the sine of the tilt angle (0.5) to obtain the penalty factor. Adding 1 and multiplying by the base of 2.8 m / s, the adaptive bias threshold is calculated to be approximately 4.1787 m / s. Since the bias of 3.0 m / s is less than the threshold, 4.1787 is added to it to obtain a dynamic denominator of 8.3574. The dynamic weight coefficient of the LSM network is calculated to be 0.5, and the dynamic weight coefficient of the Random Forest model is also 0.5. The predicted values ​​of the two models are multiplied by their respective weights and then summed to generate a fused predicted value of 15.0 m / s. This mechanism adaptively adjusts the tolerance through real terrain physical constraints, avoiding excessive weight decay under turbulent conditions on the windward slope and improving the stability of the model fusion results.

[0025] In S4, a spatial interpolation algorithm based on 3D terrain physical barrier reconstruction is used to extrapolate forecast results. The absolute vertical elevation difference and two-dimensional horizontal straight-line distance between the target location node and each wind measurement node are calculated, and these two are added together to obtain the aerodynamic equivalent distance. Assuming the target node is at an altitude of 2500m, the equivalent distance to the first wind measurement node at an altitude of 2800m and a distance of 1500m is 1800m, and its negative square is... The equivalent distance to the second wind measurement node (assumed fusion prediction value of 9.0908 m / s) at an altitude of 2400 m and a distance of 800 m is 900 m, and its negative square is... The fused predicted value of each node is multiplied by the negative square of the corresponding equivalent distance, and the sum is used as the numerator. The sum of the negative squares is used as the denominator. Dividing these results yields a wind speed forecast of approximately 10.273 m / s for the target node. This method transforms spatial elevation differences into equivalent horizontal drag, automatically weakening the wind speed correlation caused by deep valleys or ridges, eliminating extrapolation distortions when crossing terrain obstacles, and improving the accuracy and physical rationality of spatiotemporal wind speed forecasts.

[0026] In summary, the entire topographic-influence-based long short-term memory random forest wind speed prediction method firstly, by decoupling and extracting short-term fluctuation feature components within the long short-term memory network, the long-term stable meteorological background is removed. This allows the feature vectors input into the subsequent random forest model to focus on high-frequency wind speed disturbances induced by micro-topography, overcoming the smoothing failure problem of deep networks in microclimate gust prediction and enhancing the sensitivity to capturing local airflow abrupt changes. Secondly, a dynamic error allocation mechanism driven by the physical form of the data is established, replacing fixed preset parameters. By extracting the maximum difference between adjacent data in the historical wind speed sequence as the fluctuation baseline and introducing the cosine of the angle between the wind direction and the normal of the topographic slope as the trigger condition for topographic penalty, the tolerance threshold is adaptively amplified when airflow impacts the windward slope and induces mechanical turbulence. This dynamic computation mechanism considers the model prediction differences caused by local airflow fluctuations, reduces the weight decay caused by excessive judgment errors, and improves the stability of multi-model fusion results. Finally, in the spatial wind field reconstruction stage, the absolute vertical height difference between the target location node and the wind measurement node is converted into horizontal resistance, and an equivalent penalty distance is constructed. This process converts the spatial vertical height difference and horizontal distance into an equivalent distance. When two points are close in horizontal distance but have a height difference, the equivalent distance is increased and the interpolation weight is reduced, which reduces the interpolation error caused by crossing the ridge. This makes the wind speed forecast results extrapolated to the entire line fit the attenuation characteristics of the actual airflow, and provides reliable data support for power grid disaster prevention and early warning.

[0027] like Figure 2 As shown, in one specific embodiment, S1 includes: S11, acquiring the initial wind speed data and initial environmental data of each wind measurement node, and using linear interpolation to fill in the missing time node data in the initial wind speed data and initial environmental data to obtain a continuous data sequence.

[0028] S12. Subtract the minimum value in the continuous data sequence from the currently collected value, and divide by the difference between the maximum and minimum values ​​in the continuous data sequence to obtain the normalized values ​​corresponding to the historical wind speed sequence and environmental characteristic parameters.

[0029] In this embodiment, it should be noted that in S11, meteorological and topographic data are collected for subsequent wind speed prediction. Based on the rate of elevation change, multiple wind measurement nodes are deployed along the transmission line to be predicted. For example, for a target location node at an altitude of 2500m, a first wind measurement node at an altitude of 2800m is deployed on the windward side of a ridge with a drastic rate of elevation change, and a second wind measurement node at an altitude of 2400m is deployed on the leeward side of a valley. Subsequently, the data sampling frequency of the underlying sensors is set (e.g., configured to 0.1). This process involves high-frequency sampling (corresponding to 10-second intervals) to capture initial wind speed data for each anemometer node within historical timeframes, as well as initial environmental data including temperature, humidity, air pressure, and wind direction. Simultaneously, topographic feature parameters for these nodes are extracted, including elevation, slope, and orientation. This step obtains the physical parameters influencing wind speed changes, correlates meteorological observation data with three-dimensional topographic features, and increases the dimensionality of the input features.

[0030] Furthermore, missing value verification and imputation are performed on the collected time-series data to ensure the integrity of the data sequence in the time dimension. Meteorological sensors are prone to brief data loss or communication interruptions in harsh environments. First, the initial wind speed and initial environmental data are scanned to locate missing time point data. Then, linear interpolation is used to impute these missing points. For example, if the wind speed at the first wind measurement node is 14.0 m / s at 10:00 and 16.4 m / s at 10:20, but data for 10:10 is missing, linear interpolation calculates the wind speed at 10:10 to be 15.2 m / s. This process avoids the problem of the time-series model being unable to extract features due to missing data, and obtains continuous data sequences through linear interpolation, providing samples for model training.

[0031] In S13, the continuous data sequence is normalized to eliminate dimensional differences between different physical quantities. Since parameters such as wind speed, temperature, and humidity have different magnitudes and units, directly inputting them into the model would lead to difficulties in gradient descent. The minimum value in the continuous data sequence is subtracted from the currently collected value, and then divided by the difference between the maximum and minimum values. Taking the wind speed of the first wind measurement node as an example, assuming its historical maximum value is 35.0 m / s and its minimum value is 0 m / s, and the currently collected continuous sequence wind speed value is 15.2 m / s, the calculation process is as follows: The normalized value of 0.4343 corresponding to the historical wind speed sequence is obtained. Similarly, the same operation is performed on the environmental feature parameters to obtain their normalized values. Dimensionless processing brings all types of features to the same numerical order of magnitude, avoiding excessive influence of features with large values ​​on network weight updates, thereby accelerating the model's convergence speed and improving stability.

[0032] like Figure 3 As shown, in one specific implementation, S2 includes: S21, constructing a time-series feature vector based on the historical wind speed sequence, inputting the time-series feature vector into a long short-term memory network for forward propagation calculation, and obtaining the prediction value of the long short-term memory network.

[0033] S22. Extract the forget gate output vector and candidate unit state vector of the Long Short-Term Memory network at the last time step; obtain the element-wise product of the forget gate output vector and the candidate unit state vector, and determine the element-wise product as the short-term fluctuation feature component. This process isolates the stable long-term meteorological background through the low-level product and directly extracts the high-frequency wind speed disturbance features induced by the terrain.

[0034] S23. The short-term fluctuation feature components, environmental feature parameters, and terrain feature parameters are concatenated to obtain the model input vector. Multiple decision trees in the random forest model are used to calculate the model input vector, obtaining the output value of each decision tree. The arithmetic mean of all output values ​​from the multiple decision trees is obtained, and this arithmetic mean is determined as the random forest prediction value. Specifically, the number of decision trees in the random forest model is preset as follows: a typical wind speed sample set containing various terrain and meteorological conditions is obtained; an initial number of decision trees is set and gradually increased at a fixed step size; at each node, the mean squared error of the model on the cross-validation set is calculated; when the decrease in the mean squared error becomes gradual and falls below a preset convergence threshold, the number of decision trees is fixed to a preset parameter to balance prediction accuracy and computational overhead. For example, a historical sample set containing 10,000 micro-topographic wind speed data points was used for testing. The initial number of trees was set to 10, and the step size was set to 10. After testing, when the number of trees increased to 100, the mean square error of the validation set decreased from 2.5 to 0.8. When the number of trees was further increased to 110, the error only decreased slightly to 0.79, with a very small convergence amplitude. Based on this, the number of decision trees in the random forest was fixed and preset to 100.

[0035] In this embodiment, it should be noted that in S21, normalized data is used to construct time series samples suitable for deep learning. A preset historical time window length is obtained, and the instantaneous wind speed data of the wind measurement node corresponding to the prediction time is arranged chronologically with the wind speed data of the previous five historical times to form a time series feature vector. The specific process of obtaining the preset historical time window length is as follows: extract the historical wind speed sequence of the wind measurement node over the past year, and calculate the autocorrelation coefficient of this data sequence at different time lag steps. As the lag step increases, when the autocorrelation coefficient first decays and falls below a set correlation threshold, this lag step number is determined as the preset historical time window length to balance evolutionary information and noise. For example, when calculating the autocorrelation coefficient of the historical wind speed sequence of a certain wind measurement node, with a correlation threshold of 0.3, the coefficient is greater than 0.3 when the lag step is 1 to 4, and decays to 0.28 (below 0.3) in step 5. Based on this, the preset historical time window length is determined to be 5 historical times.

[0036] Subsequently, the temporal feature vector is input into a Long Short-Term Memory (LSTM) network for forward propagation calculation. LTM, with its gating mechanism, can memorize historical evolution patterns in the wind speed sequence. Taking the first wind measurement node mentioned above as an example, its temporal feature vector, containing 15.2 m / s (normalized to 0.4343) and historical time data, is input into the network. The network iteratively updates the hidden layer states and finally obtains the LTM prediction value at the output layer. Assuming the LTM prediction value for this node is 16.5 m / s, this step leverages the advantage of deep learning networks in handling time dependencies to extract the macroscopic evolution trend of the wind speed sequence, providing a preliminary time-dimensional prediction benchmark for fusion prediction.

[0037] In S22, asymmetric feature decoupling is performed within the network to extract key short-term fluctuation information. Existing fusion models typically use predicted outputs or the full hidden layer states directly, which can easily be mixed in with long-term stable weather backgrounds. The forget gate output vector and candidate unit state vector of the Long Short-Term Memory network at the last time step are extracted. The forget gate controls the proportion of historical information discarded, and the candidate units contain newly added input features at the current time step. The element-wise product of these two vectors is obtained and determined as the short-term fluctuation feature component. Specifically, the element-wise product expression can be represented as... ;in, This represents the short-term fluctuation characteristic component of the k-th wind measurement node at the t-th prediction time; This represents the forget gate output vector of the Long Short-Term Memory network at time step t. This represents the state vector of a candidate unit in a Long Short-Term Memory network at time step t. This represents the element-wise multiplication operation of vectors.

[0038] For example, suppose the dimension of a certain feature in the forget gate output vector is 0.85, and the corresponding dimension of the candidate unit state vector is 0.60; the element-wise product result is 0.51. This specific low-level multiplication operation mathematically eliminates long-term stationary features that the network determines should be forgotten, while selectively retaining high-frequency wind speed abrupt changes induced by the current micro-topography. This processing achieves decoupling of the temporal model state, providing perturbation feature input for subsequent nonlinear mapping.

[0039] In step S23, the decoupled temporal features and spatial physical features are subjected to multi-source cascade mapping. Short-term fluctuation feature components (e.g., 0.51), environmental feature parameters, and terrain feature parameters are concatenated to obtain the model input vector. This vector integrates recent wind speed disturbances, real-time meteorological conditions, and fixed terrain constraints. Subsequently, multiple decision trees in the random forest model are used to calculate the model input vector, obtaining the output value of each decision tree. Finally, the arithmetic mean of all output values ​​from multiple decision trees is obtained, and this result is determined as the random forest prediction value. For example, for the first wind measurement node, the arithmetic mean calculated by multiple decision trees, i.e., the random forest prediction value, is 13.5 m / s. This step leverages the advantage of the random forest model in handling high-dimensional nonlinear relationships, cross-mapping temporal fluctuation features with static terrain and dynamic environmental features, overcoming the limitation of a single deep learning model in handling spatial heterogeneity, and outputting independent prediction reference values ​​specific to the terrain and environment.

[0040] like Figure 4 As shown, in one specific implementation, S3 includes: S31, obtaining the absolute value of the difference between the prediction value of the Long Short-Term Memory Network and the prediction value of the Random Forest, and using this absolute value as the model bias; extracting the maximum value of the absolute value of the difference between adjacent data in the historical wind speed sequence, and using the maximum value as the tolerance threshold base. Here, the tolerance threshold base represents the natural fluctuation extreme value of the wind measurement node in the recent real wind field, providing an objective data baseline for subsequent calculations; obtaining the measured wind direction angle of the wind measurement node, as well as the normal azimuth angle and tilt angle in the terrain feature parameters.

[0041] S32. When the cosine of the difference between the measured wind direction angle and the normal azimuth angle is positive, the cosine value is multiplied by the sine of the tilt angle to obtain the penalty multiplier. Here, the penalty multiplier is extracted as a numerical multiplier from the physical shearing effect of the terrain slope on the airflow, and is triggered only when the airflow directly hits the windward slope; when the cosine value is not positive, the penalty multiplier is set to zero; the numerical value is added to the penalty multiplier to obtain the scaling factor, and the tolerance threshold base is multiplied by the scaling factor to obtain the adaptive deviation threshold. The calculation expression for the adaptive deviation threshold is:

[0042] in, This represents the adaptive deviation threshold of the k-th wind measurement node at the t-th prediction time; This represents the wind speed data of the k-th wind measurement node at historical time tj. This represents the wind speed data of the k-th wind measurement node at historical time tj-1; This represents the measured wind direction angle at the k-th wind measurement node at the t-th prediction time; This represents the normal azimuth angle in the terrain feature parameters of the k-th wind measurement node; The tilt angle represents the topographic feature parameter of the k-th wind measurement node; The index number representing the historical time step, with a value range of 1. to .

[0043] S33. Obtain the maximum value between the adaptive bias threshold and the model bias, and add the maximum value to the adaptive bias threshold to obtain the dynamic denominator. Here, the dynamic denominator is the dynamic calculation base used for normalization allocation, which ensures that the weights decay smoothly as the prediction error increases. Divide the adaptive bias threshold by the dynamic denominator to obtain the first dynamic weight coefficient of the Long Short-Term Memory network. Subtract the first dynamic weight coefficient from the value to obtain the second dynamic weight coefficient of the Random Forest model. Multiply the predicted value of the Long Short-Term Memory network by the first dynamic weight coefficient to obtain the first product, multiply the predicted value of the Random Forest by the second dynamic weight coefficient to obtain the second product, and add the first product and the second product to obtain the fused predicted value.

[0044] In this embodiment, it should be noted that in S31, the data basis evaluation for initiating the error allocation mechanism is performed. The absolute value of the difference between the Long Short-Term Memory network prediction (16.5 m / s) and the Random Forest prediction (13.5 m / s) is obtained, and the result of 3.0 m / s is used as the model bias. Next, the maximum absolute value of the difference between adjacent data in the historical wind speed sequence is extracted. Assuming that the adjacent wind speed of the first wind measurement node changed from 13.2 m / s to 16.0 m / s, the maximum difference is 2.8 m / s, and this maximum value is used as the tolerance threshold base. Subsequently, the measured wind direction angle (270°) of the wind measurement node, as well as the normal azimuth angle (280°) and tilt angle (30°) in the terrain feature parameters, are obtained. This step does not use a pre-set fixed error standard, but uses the recent real physical fluctuation amplitude of the wind measurement node (2.8 m / s) as the evaluation baseline, and simultaneously obtains the physical parameters required to determine the interaction angle between airflow and terrain, providing objective and real-time data input for subsequent adaptive calculation of the terrain penalty limit.

[0045] In S32, the adaptive deviation threshold is calculated by combining the physical angle. Specifically, the wind-slope interaction mechanism in the underlying fluid dynamics is transformed into specific tolerance constraint values ​​to establish an objective data-driven evaluation baseline. Its complete calculation expression is as follows: .

[0046] The first part on the left of the expression The aim is to extract the maximum absolute value of the difference between adjacent data in historical wind speed sequences. This calculation obtains the maximum value of the wind speed difference within the past five time steps, extracting the maximum disturbance amplitude of the recent wind speed. By using this value as the tolerance threshold base instead of a fixed preset error constant, the evaluation benchmark of the model can be anchored in real time to the actual natural turbulence level at that spatial location. For example, in a specific application scenario, it was found that the wind speed at this node jumped from 13.2 m / s to 16.0 m / s, thus extracting the maximum absolute value of the difference, 2.8 m / s, as the basic tolerance boundary.

[0047] Furthermore, the bracketed portion on the right side of the expression constitutes a dimensionless terrain penalty multiplier, used to quantify the intensity of mechanical turbulence generated when a specific wind direction interacts with terrain geometry.

[0048] in, The calculation is the angle between the measured wind direction angle at the predicted time and the azimuth angle of the slope normal at the wind measurement node, and its corresponding cosine function. Physically, the calculation represents the orthogonal projection component of the wind speed vector onto the windward surface of the terrain. When the wind direction hits the slope perpendicularly, the cosine value approaches one, indicating that the horizontal kinetic energy of the airflow is converted into vertical shear energy. Based on specific data, when the measured wind direction angle is 270° and the normal azimuth angle is 280°, the difference between the two is -10°, and the cosine value is approximately 0.9848. This indicates that the airflow is forming a high-intensity frontal impact on the slope.

[0049] at the same time, outer layer In mathematics, the function acts as a nonlinear rectifier, which establishes a strict directional filtering mechanism: when the wind blows from the leeward side (i.e., the angle between the wind direction and the normal exceeds 90 degrees and the cosine value is negative), the airflow separates instead of being forced to rise. At this time, the function forces the output to be zero, thereby neutralizing the inapplicable mechanical turbulence penalty; while in the current upwind scenario, the function extracts and retains the positive value of 0.9848.

[0050] Next, the projection component The sine value of the slope angle relative to the terrain Multiply, because the steeper the slope (the larger the sine value), the more intense the forced lifting effect caused by the airflow encountering resistance and the separation phenomenon from the lower boundary layer; substituting the tilt angle of 30°, its sine value is 0.5, multiplying it by 0.9848 gives the value 0.4924, and this product calculates the topographic penalty multiple for a specific slope and wind direction.

[0051] Finally, by adding a constant of 1, a scaling factor of 1.4924 is obtained, ensuring that even in flat terrain or leeward environments (where the penalty multiplier is zero), the overall multiplier remains at least at the base multiplier of 1, preventing the base tolerance from being reset to zero. Multiplying the tolerance threshold base of 2.8 m / s by the scaling factor 1.4924, the adaptive deviation threshold for this node is finally calculated to be approximately 4.1787 m / s.

[0052] This calculation process directly solves the technical problem of weight erratic attenuation caused by static error standards in multi-model fusion. When actual airflow impacts steep mountains and causes strong wind field distortion, different prediction models (such as memory networks that focus on time series and random forests that focus on spatial mapping) will inevitably produce model biases such as 3.0 m / s due to their different feature sensitivities. If the terrain coupling expression is not introduced, the 3.0 m / s bias may be judged as a prediction failure and the model weights may be reduced. Through the calculation of this expression, it is confirmed that the tolerance threshold of 4.1787 m / s under the current wind direction and slope conditions is within the fluctuation range caused by normal physical turbulence. Therefore, the 3.0 m / s bias is accepted as a reasonable fluctuation under terrain disturbance, thereby maintaining the dynamic balance of model weights and ensuring the rationality of the fusion algorithm under extreme airflow impact.

[0053] In S33, the weights of the two models are dynamically allocated based on the deviation threshold. For the first wind measurement node, the model deviation of 3.0 m / s is less than the adaptive deviation threshold of 4.1787 m / s. The maximum value of the two, 4.1787, is obtained and added to the adaptive deviation threshold of 4.1787 to obtain a dynamic denominator of 8.3574. The adaptive deviation threshold is divided by the dynamic denominator to obtain the first dynamic weight coefficient of the Long Short-Term Memory network, which is 0.5. The first dynamic weight coefficient is subtracted from the value to obtain the second dynamic weight coefficient of the Random Forest model, which is also 0.5. The prediction value of the Long Short-Term Memory network (16.5 m / s) is multiplied by the first dynamic weight coefficient (0.5) to obtain the first product (8.25), and the prediction value of the Random Forest (13.5 m / s) is multiplied by the second dynamic weight coefficient (0.5) to obtain the second product (6.75). The first product and the second product are added to obtain the fused prediction value of 15.0 m / s. This computational mechanism ensures that balanced weights are assigned when the predicted differences match the expected topographic turbulence, while when the error surges abnormally (such as a deviation of 4.0 m / s on the leeward side exceeding the threshold of 1.5 m / s), it drives the weights to decay towards the robust model, thus improving the reliability of the fusion results.

[0054] like Figure 5As shown, in one specific embodiment, S4 includes: S41, obtaining the two-dimensional horizontal straight-line distance between the target location node and the wind measurement node, and the absolute vertical height difference between the altitude of the target location node and the altitude of the wind measurement node; adding the two-dimensional horizontal straight-line distance and the absolute vertical height difference to obtain the aerodynamic equivalent distance. This conversion overcomes the defect of existing models that ignore mountain obstruction during spatial interpolation.

[0055] S42. Multiply the fused predicted value of each wind measurement node by the negative square of the corresponding aerodynamic equivalent distance and sum them to obtain the interpolation numerator; sum the negative squares of the aerodynamic equivalent distances of each wind measurement node to obtain the interpolation denominator; divide the interpolation numerator by the interpolation denominator to obtain the wind speed forecast result for the target location node. The calculation expression for the wind speed forecast result is as follows:

[0056] in, This represents the wind speed forecast result for the target location node p at the t-th prediction time. This indicates the total number of wind measurement nodes involved in the calculation; This represents the fused prediction value of the k-th wind measurement node at the t-th prediction time; This represents the two-dimensional horizontal straight-line distance between the target location node p and the k-th wind measuring node; This represents the altitude of the target location node p; This represents the altitude of the k-th wind measurement node.

[0057] S43. When there are multiple target location nodes and they are distributed in the target space grid, extract the maximum wind speed value and the average wind speed value from the wind speed forecast results of all target location nodes at the current prediction time; continuously extract the maximum wind speed value and the average wind speed value at multiple prediction times to generate a wind speed forecast time series curve.

[0058] S44. Compare the wind speed forecast time series curve with the pre-stored safe extreme value range. When the maximum wind speed value is outside the safe extreme value range, generate wind speed warning information including spatial coordinate range and expected occurrence time. The specific generation process of the pre-stored safe extreme value range is as follows: Extract the critical values ​​of disaster-causing wind speeds that lead to conductor galloping or tower deformation from the historical mechanical disaster records of the target line. Compare these with the maximum wind resistance standard limit of the power grid engineering design specifications for the area. Take the minimum value of the two and deduct the disaster prevention safety margin to generate the safe extreme value range. For example, if the historical archives of a certain line show that the lowest recorded wind speed causing conductor galloping is 26.5 m / s, while the designed wind resistance standard for this section is 28.0 m / s, the algorithm extracts the smaller value of 26.5 m / s. To reserve disaster prevention response time, a safety margin of 1.5 m / s is deducted, and the upper limit is defined as 25.0 m / s, thus generating a pre-stored safe extreme value range from 0 to 25.0 m / s.

[0059] In this embodiment, it should be noted that in S41, the distance attenuation metric for reconstructing spatial interpolation incorporates terrain elevation differences into the calculation. The two-dimensional horizontal straight-line distance between the target location node (elevation 2500m) and the wind measurement node, as well as their absolute vertical elevation difference, are obtained. For the first wind measurement node at an altitude of 2800m and a distance of 1500m, the absolute vertical elevation difference is 300m; for the second wind measurement node at an altitude of 2400m and a distance of 800m, the absolute vertical elevation difference is 100m. The two-dimensional horizontal straight-line distance and the absolute vertical elevation difference are added together to obtain the aerodynamic equivalent distance. The equivalent distance for the first wind measurement node is 1800m, and the equivalent distance for the second wind measurement node is 900m. This step breaks the limitation of existing algorithms that rely solely on planar distance. Its physical logic lies in materializing the mountain obstruction and valley elevation difference in three-dimensional space as horizontal resistance to airflow spread. By increasing the equivalent distance, the airflow isolation effect caused by elevation differences can be quantified, providing a distance benchmark that conforms to aerodynamic attenuation characteristics for subsequent weight allocation.

[0060] In S42, spatial interpolation is performed based on the equivalent distance to output the target forecast. Specifically, after data fusion at local nodes is completed, the discrete forecast data is extended to the entire target area using a specific spatial interpolation expression. The calculation expression for the core wind speed forecast result is as follows: The numerator and denominator of this expression together form a weighted average framework based on aerodynamic equivalent distance, which is used to solve the spatial mapping error problem caused by the failure to consider three-dimensional physical barriers in complex mountain microclimates in existing spatial interpolation algorithms.

[0061] Furthermore, the core of the expression lies in the base part. The existing inverse distance weighting method relies solely on two-dimensional horizontal straight-line distance. This algorithm assumes that the kinetic energy of the wind field decays uniformly in planar space, without considering the hydrodynamic characteristics of airflow being blocked by ridges or deep valleys in canyons and mountains. To correct this bias, the expression introduces the absolute vertical height difference between the target location node and the wind measurement node. This is then combined with the two-dimensional horizontal distance as a direct additive term. The physical significance of this mathematical combination lies in transforming the vertical terrain obstruction effect into the horizontal aerodynamic equivalent extension drag, because in the real atmospheric boundary layer, the kinetic energy consumed by the airflow to overcome gravity and climb over the elevation difference is far greater than the kinetic energy loss when it glides across the same horizontal plane.

[0062] Based on data from specific application scenarios, it is necessary to extrapolate the altitude. For the wind speed at the target location node at 2500m, and for the first wind measurement node, its fused prediction value is... The speed is 15.0 m / s, and the altitude is [missing information]. The distance is 2800m, a two-dimensional horizontal straight-line distance. It is 1500m; its absolute vertical height difference is Adding 1500m and 300m together yields the aerodynamically equivalent distance of 1800m. Subsequently, the expression is calculated using a negative square. Right now A distance decay weight was constructed, and this nonlinear exponential decay indicates that an increase in the equivalent distance leads to a significant decrease in the influence weight; 15.0 m / s and Multiplication results in an additive term in the numerator representing the influence of that node. Similarly, for the fusion prediction value... The second wind measurement node has a wind speed of 9.0908 m / s and an altitude of [altitude missing]. It is 2400m, horizontal distance Given a height of 800m, calculate the absolute vertical height difference as follows: The combined equivalent distance is 900m, and its weighting factor is... Right now Multiplying this by 9.0908 m / s forms the second numerator summation term. The denominator of the expression is the sum of these two weighting factors. and A simple summation is performed to apply normalized division to the final result. The summation of the numerator terms is divided by the summation of the denominator terms, resulting in a final wind speed forecast of approximately 10.273 m / s for the target location node.

[0063] This detailed division process clearly demonstrates how the expression fundamentally solves the problem of inference distortion when crossing terrain obstacles: Although the first wind measurement node measured a relatively high wind speed of 15.0 m / s, it is located on a 2800m mountain ridge, with a vertical physical difference of up to 300m from the target node. This huge elevation misalignment indicates that the two are in mutually isolated local wind field streamlines; the expression reduces the influence weight of the higher node on the calculation results of the lower node by directly superimposing the 300m elevation difference to the horizontal distance and performing negative square attenuation. Conversely, although the second wind measurement node has a weaker wind speed of only 9.0908 m / s, it is not only closer in horizontal distance, but also has a small elevation difference of 100 m, which puts it in a more coherent airflow channel with the target node in terms of aerodynamics. Therefore, the expression gives it a dominant interpolation weight through a smaller equivalent distance, making the forecast result of 10.273 m / s consistent with the attenuation law of valley airflow, reducing the error caused by the downward interpolation of strong wind data at high altitudes, and ensuring the objectivity and accuracy of the spatial gridded forecast results.

[0064] In S43, spatial interpolation results are processed to analyze time-series forecast trends. When multiple target location nodes exist and are distributed across the target spatial grid, the maximum and average wind speed values ​​from the wind speed forecasts of all target location nodes at the current forecast time are extracted. The specific construction process of the target spatial grid is as follows: Digital elevation model data of the transmission line area to be predicted is acquired, and the three-dimensional spatial boundary enveloping the target geographical area is extracted. The grid spatial resolution is set according to the wind field forecast accuracy requirements. This resolution is used to discretize the spatial boundary, and the coordinates of each grid point are extracted to generate the target spatial grid. For example, the geographical envelope boundary of a cross-valley transmission corridor (10 km long, 5 km wide) is extracted, and the spatial resolution is set to 100 meters. After two-dimensional discretization of the boundary, 100 and 50 nodes are generated in the length and width directions, respectively. Each node is matched with an elevation, ultimately constructing a target spatial grid with 5000 nodes and a 100-meter resolution.

[0065] For example, during a certain forecast period, the maximum wind speed forecast for multiple nodes along the target transmission line is 10.273 m / s, and the average wind speed is 8.5 m / s. Subsequently, the maximum and average wind speed values ​​at multiple forecast times are continuously extracted and a time-series wind speed forecast curve is generated in chronological order. This step reorganizes the discrete spatial forecast results along the time axis, presenting the evolution pattern and extreme value fluctuation trend of the wind field along the transmission line, providing structured time-series state data support for subsequent safety assessments and early warning triggering.

[0066] In step S44, disaster prevention and early warning judgment is performed based on the wind speed forecast time series curve. The wind speed forecast time series curve is compared with the pre-stored safe extreme value range. It is assumed that the upper limit of the safe extreme value range for transmission lines is 25.0 m / s. When the wind speed forecast time series curve shows that the maximum wind speed value (e.g., reaching 28.0 m / s) is outside the safe extreme value range, a response mechanism is triggered to generate wind speed early warning information containing the spatial coordinate range and the expected occurrence time. This early warning information indicates the specific section and time period where the expected super-strong winds will occur. This step transforms the high-precision spatiotemporal wind speed prediction results into direct engineering safety instructions, enabling power grid dispatching and operation and maintenance departments to obtain reliable data forecasts before disastrous microclimate events occur, thereby taking timely wind prevention and disaster reduction measures and ensuring the safe operation of transmission lines.

[0067] This invention also provides a long short-term memory random forest wind speed prediction system based on terrain influence. The system is used to implement a long short-term memory random forest wind speed prediction method based on terrain influence. The system includes: The parameter acquisition module is used to acquire historical wind speed sequences, environmental characteristic parameters, and terrain characteristic parameters for each wind measurement node, and perform normalization processing. The feature decoupling module is used to construct a time-series feature vector based on historical wind speed sequences, input the time-series feature vector into a long short-term memory network, obtain the prediction value of the long short-term memory network, and extract the short-term fluctuation feature components inside the long short-term memory network. The model mapping module is used to concatenate short-term fluctuation feature components, environmental feature parameters, and terrain feature parameters to obtain the model input vector, and then input the model input vector into the random forest model to obtain the random forest prediction value. The error allocation module is used to obtain the model bias based on the prediction values ​​of the long short-term memory network and the random forest, obtain the adaptive bias threshold based on the difference between adjacent data of historical wind speed sequences and terrain feature parameters, obtain the dynamic weight coefficient based on the adaptive bias threshold and the model bias, and generate the fusion prediction value of the wind measurement node based on the dynamic weight coefficient. The spatial reconstruction module is used to obtain the aerodynamic equivalent distance based on the vertical height difference and horizontal distance between the target location node and the wind measurement node, and to perform spatial interpolation processing on the fused prediction value based on the aerodynamic equivalent distance to obtain the wind speed forecast result of the target location node.

[0068] In one specific implementation, when the error allocation module obtains the adaptive deviation threshold based on the difference between adjacent data in the historical wind speed sequence and the terrain feature parameters, it specifically performs the following steps: extracting the maximum absolute value of the difference between adjacent data in the historical wind speed sequence and using the maximum value as the tolerance threshold base; obtaining the measured wind direction angle of the wind measurement node, as well as the normal azimuth angle and tilt angle in the terrain feature parameters; when the cosine value of the difference between the measured wind direction angle and the normal azimuth angle is positive, multiplying the cosine value by the sine value of the tilt angle to obtain the penalty multiple; when the cosine value is not positive, setting the penalty multiple to zero; adding the value one to the penalty multiple to obtain the scaling factor, and multiplying the tolerance threshold base by the scaling factor to obtain the adaptive deviation threshold.

[0069] To further clarify the operating mechanism and physical quantification process of the technical solution of this invention, the following analysis will be conducted in detail on the underlying derivation logic of the long short-term memory random forest wind speed prediction method and system based on terrain influence, using a scenario containing specific parameters and data.

[0070] Taking a high-altitude canyon wind farm as an example, a target location node with an altitude of 2500m is set along the transmission line to be predicted. Two wind measurement nodes are deployed around the target location node. The first wind measurement node is located on the windward side of the ridge, at an altitude of 2800m, and is 1500m away from the target location node in a two-dimensional horizontal straight line. The second wind measurement node is located on the leeward side of the canyon, at an altitude of 2400m, and is 800m away from the target location node. In S1, the system first obtains the wind speed sequence of the first wind measurement node over a historical period, assuming that its historical maximum wind speed is 35.0m / s, the minimum wind speed is 0m / s, and the currently collected real-time wind speed is 15.2m / s. After the system fills in the abnormal gaps using linear interpolation, it performs normalization processing and calculates the normalized value of the current wind speed as 0.4343. This step effectively eliminates the dimensional differences caused by different meteorological sensor ranges, providing a standard data stream for the stable convergence of the subsequent deep learning network.

[0071] In S2, the system constructs a time-series feature vector using historical normalized data and inputs it into a pre-trained Long Short-Term Memory (LSTM) network. The specific training process for the pre-trained LSM network and the aforementioned random forest model is as follows: Synchronous high-frequency historical monitoring data from multiple meteorological stations in the target area over the past few years are collected. After outlier removal and the aforementioned normalization process, a large-scale supervised learning sample library is constructed. The sample library is divided into a training set and a validation set according to time series. Using the root mean square error as the loss function, an adaptive moment estimation optimization algorithm is employed to perform multiple rounds of backpropagation on the LSM network to iteratively update the weights until the validation set loss converges. Simultaneously, a decision tree is constructed on the training set using a feature bootstrapping method to complete the training and solidification of the random forest model. For example, 100,000 labeled time-series wind speed and terrain correlation data from the past three years were extracted from a proposed wind farm. The training set and validation set were divided into an 8:2 ratio. The initial learning rate was set to 0.001. After 50 full-sample iteration cycles, the root mean square error of the validation set of the Long Short-Term Memory Network stabilized at 0.45 m / s and met the requirements, thus solidifying the weight parameters of the network nodes inside the model in advance.

[0072] For the first wind measurement node, the predicted value output by the Long Short-Term Memory (LSTM) network after forward propagation is 16.5 m / s. Simultaneously, to avoid introducing long-term stable meteorological characteristics into the nonlinear mapping process, the system decouples and extracts the short-term fluctuation feature components from the network's underlying layers. In the final time step, assuming that one feature dimension of the extracted forget gate output vector has a value of 0.85 and the corresponding value of the candidate unit state vector is 0.60, the system calculates the element-wise product expression... The short-term fluctuation feature component of this dimension was obtained as 0.51. This processing stripped away the macro-climate background and extracted the high-frequency turbulent disturbance features caused by the mountain micro-topography. Then, this component was concatenated with environmental features and topographic features and input into the random forest model. The output random forest prediction value of the first wind measurement node was 13.5 m / s.

[0073] Upon entering S3, the system initiates an error allocation mechanism based on physical terrain constraints. First, it calculates the model deviation for the first wind measurement node, i.e., an absolute difference of 3.0 m / s. Then, it calculates the adaptive deviation threshold. The system traces the wind speed data of this node over the past five moments and finds that the largest adjacent wind speed abrupt change occurred two minutes prior, with the wind speed jumping from 13.2 m / s to 16.0 m / s. Therefore, the maximum absolute value of the difference, 2.8 m / s, is extracted as the tolerance threshold base. The meteorological sensor simultaneously measures a real-time wind direction angle of 270°, while the normal azimuth angle of the slope where the first wind measurement node is located is 280°, and the tilt angle is 30°. Since the airflow is almost directly hitting the slope at this time, the difference between the two angles is -10°, and the sine value is approximately 0.9848, which is greater than 0; simultaneously, the sine value of the tilt angle is equal to 0.5. The system calculates the adaptive deviation threshold according to the expression: The calculation results show that the adaptive bias threshold of this node is 4.1787 m / s. This value intuitively transforms the strong mechanical turbulence generated by the windward airflow rise into an amplified model tolerance boundary.

[0074] After determining the adaptive bias threshold, S3 continues to calculate the dynamic weight coefficients. For the first wind measurement node, its model bias of 3.0 m / s is less than the adaptive bias threshold of 4.1787 m / s, indicating that the difference in predictions between the two models falls within the normal fluctuation range under complex turbulence on the windward slope. The system takes the maximum value of the two, 4.1787, and adds it to the adaptive bias threshold to obtain a dynamic denominator of 8.3574. Dividing 4.1787 by 8.3574, the dynamic weight coefficient of the Long Short-Term Memory network is calculated to be 0.5. Subtracting 0.5 from the value 1 yields a dynamic weight coefficient of 0.5 for the Random Forest model. Finally, the system generates the fused prediction value for the first wind measurement node through weighted summation: 0.5 × 16.5 + 0.5 × 13.5 = 15.0 m / s.

[0075] In contrast, the second wind measurement node at the same time, being on the leeward slope, has a negative cosine value for the difference between its measured wind direction and the normal azimuth angle, and the system forces its penalty multiplier to zero. At this point, its tolerance threshold base becomes its adaptive bias threshold, assumed to be 1.5 m / s. The Long Short-Term Memory (LSTM) network predicts 12.0 m / s, and the Random Forest predicts 8.0 m / s, resulting in a model bias as high as 4.0 m / s. This bias far exceeds the threshold, reaching a maximum of 4.0. Adding the threshold of 1.5 yields a dynamic denominator of 5.5. The LSM weights decay to 1.5 divided by 5.5, approximately 0.2727, while the Random Forest weights are passively increased to 0.7273. The final fused prediction value for the second wind measurement node is: 0.2727 × 12.0 + 0.7273 × 8.0 = 9.0908 m / s. This mechanism adaptively adjusts the model weights based on data characteristics, reducing the risk of prediction failure due to fixed constants.

[0076] In S4, the system needs to project the predicted values ​​of discrete nodes to the target location node. Existing algorithms only consider two-dimensional distances, while this scheme constructs an aerodynamic equivalent distance based on the vertical elevation difference and horizontal distance. For the first wind-measuring node, its absolute vertical elevation difference with the target location node is 300m, and its horizontal distance is 1500m. Adding these two directly yields an equivalent distance of 1800m. The negative square of this aerodynamic equivalent distance is... Similarly, the elevation difference of the second wind measurement node is 100m, the horizontal distance is 800m, and the equivalent distance is 900m, which is the negative square of... The system uses expressions for spatial interpolation: The final calculated wind speed forecast for the target location node is 10.273 m / s.

[0077] This spatial reconstruction logic physically transforms the elevation difference of the three-dimensional terrain into the equivalent horizontal resistance of airflow spread. Although the wind speed at the first wind measurement node is relatively high, its location on a high ridge with an elevation difference of 300m results in significant resistance to the descending cold air reaching the target node, thus reducing its interpolation weight. Conversely, the second wind measurement node, which is closer and has an elevation difference of only 100m, receives a weight contribution as high as 80%, making the forecast result of 10.273m / s more consistent with the actual wind field attenuation characteristics of the leeward valley. Finally, the system generates a wind speed forecast time series curve from the continuously acquired forecast values. If the wind speed exceeds the safe extreme range of 25.0m / s, a wind speed warning is immediately generated. This application scenario verifies the process by which the scheme transforms mechanical turbulence and terrain impedance mechanisms into mathematical boundaries.

[0078] The preferred embodiments of this disclosure have been described in detail above with reference to the accompanying drawings. However, this disclosure is not limited to the specific details of the above embodiments. Within the scope of the technical concept of this disclosure, various simple modifications can be made to the technical solutions of this disclosure, and these simple modifications all fall within the protection scope of this disclosure.

[0079] It should also be noted that the various specific technical features described in the above embodiments can be combined in any suitable manner without contradiction. To avoid unnecessary repetition, this disclosure will not describe the various possible combinations separately.

[0080] Furthermore, various different embodiments of this disclosure can be combined in any way, as long as they do not violate the spirit of this disclosure, they should also be regarded as the content disclosed in this disclosure.

[0081] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention, and they should all be covered within the scope of the claims and specification of the present invention.

Claims

1. A method for predicting wind speed in long short-term memory stochastic forests based on topographic influence, characterized in that the method... include: Historical wind speed sequences, environmental characteristic parameters, and terrain characteristic parameters of each wind measurement node are obtained and normalized. A time-series feature vector is constructed based on historical wind speed sequences. The time-series feature vector is then input into a long short-term memory network (LSTM) to obtain the LTM prediction value. Short-term fluctuation feature components within the LTM are then extracted. The short-term fluctuation feature components, environmental feature parameters, and terrain feature parameters are concatenated to obtain the model input vector, and the model input vector is input into the random forest model to obtain the random forest prediction value. The model bias is obtained by comparing the predictions from the Long Short-Term Memory Network and the Random Forest, and an adaptive bias threshold is obtained based on the difference between adjacent data in the historical wind speed sequence and terrain feature parameters. Dynamic weight coefficients are obtained based on adaptive bias threshold and model bias, and fused prediction values ​​of wind measurement nodes are generated based on dynamic weight coefficients. Based on the vertical height difference and horizontal distance between the target location node and the wind measurement node, the aerodynamic equivalent distance is obtained. The fused prediction value is then spatially interpolated based on the aerodynamic equivalent distance to obtain the wind speed forecast result for the target location node.

2. The method for predicting wind speed in a long short-term memory stochastic forest based on topographic influence as described in claim 1, characterized in that, The extraction of short-term fluctuation feature components within the Long Short-Term Memory (LSTM) network includes: Extract the forget gate output vector and candidate unit state vector of the Long Short-Term Memory network at the last time step; Obtain the element-wise product of the forget gate output vector and the candidate unit state vector, and determine the element-wise product as the short-term fluctuation feature component.

3. The method for predicting wind speed in a long short-term memory stochastic forest based on topographic influence as described in claim 1, characterized in that, The step of inputting the model input vector into the random forest model to obtain the random forest prediction value includes: The output value of each decision tree is obtained by calculating the model input vector using multiple decision trees in the random forest model. Obtain the arithmetic mean of all output values ​​from multiple decision trees, and use the arithmetic mean as the predicted value for the random forest.

4. The method for predicting wind speed in a long short-term memory stochastic forest based on topographic influence as described in claim 1, characterized in that, The method for obtaining an adaptive deviation threshold based on the difference between adjacent data in historical wind speed sequences and terrain feature parameters includes: Extract the maximum absolute value of the difference between adjacent data in the historical wind speed sequence, and use the maximum value as the tolerance threshold base. Obtain the measured wind direction angle at the wind measurement node, as well as the normal azimuth and tilt angle in the terrain feature parameters; When the cosine of the difference between the measured wind direction angle and the normal azimuth angle is positive, the cosine value is multiplied by the sine of the tilt angle to obtain the penalty multiplier; when the cosine value is not positive, the penalty multiplier is set to zero. Add the value to the penalty factor to obtain the scaling factor, and multiply the tolerance threshold base by the scaling factor to obtain the adaptive deviation threshold.

5. The method for predicting wind speed in a long short-term memory stochastic forest based on topographic influence as described in claim 1, characterized in that, The step of obtaining dynamic weight coefficients based on adaptive deviation thresholds and model deviations, and generating fused prediction values ​​for wind measurement nodes based on these dynamic weight coefficients, includes: Obtain the maximum value between the adaptive bias threshold and the model bias, and add the maximum value to the adaptive bias threshold to obtain the dynamic denominator; Divide the adaptive bias threshold by the dynamic denominator to obtain the first dynamic weight coefficient of the long short-term memory network. Subtracting the first dynamic weight coefficient from the value one yields the second dynamic weight coefficient of the random forest model; The first product is obtained by multiplying the prediction value of the Long Short-Term Memory Network by the first dynamic weight coefficient, and the second product is obtained by multiplying the prediction value of the Random Forest by the second dynamic weight coefficient. The first product and the second product are added together to obtain the fused prediction value.

6. The method for predicting wind speed in long short-term memory stochastic forests based on topographic influence as described in claim 1, characterized in that, The process involves obtaining the aerodynamic equivalent distance based on the vertical elevation difference and horizontal distance between the target location node and the wind measurement node, and then performing spatial interpolation processing on the fused prediction value based on the aerodynamic equivalent distance to obtain the wind speed forecast result for the target location node, including: Obtain the two-dimensional horizontal straight-line distance between the target location node and the wind measurement node, as well as the absolute vertical height difference between the altitude of the target location node and the altitude of the wind measurement node; Adding the two-dimensional horizontal straight-line distance to the absolute vertical height difference yields the aerodynamic equivalent distance; The interpolation numerator is obtained by multiplying the fused predicted value of each wind measurement node by the negative square of the corresponding aerodynamic equivalent distance and summing the results. The negative square of the aerodynamic equivalent distance of each wind measurement node is summed to obtain the interpolation denominator. Divide the interpolation numerator by the interpolation denominator to obtain the wind speed forecast result for the target location node.

7. The method for predicting wind speed in a long short-term memory stochastic forest based on topographic influence as described in claim 1, characterized in that, The process of acquiring historical wind speed sequences, environmental characteristic parameters, and terrain characteristic parameters for each wind measurement node, and performing normalization processing, includes: The initial wind speed data and initial environmental data of each wind measurement node are obtained. The missing time node data in the initial wind speed data and initial environmental data are filled in by linear interpolation to obtain a continuous data sequence. The normalized values ​​corresponding to the historical wind speed sequence and environmental characteristic parameters are obtained by subtracting the minimum value in the continuous data sequence from the currently collected value and dividing by the difference between the maximum and minimum values ​​in the continuous data sequence.

8. The method for predicting wind speed in a long short-term memory stochastic forest based on topographic influence as described in claim 1, characterized in that, After obtaining the wind speed forecast results for the target location node, the following is also included: When there are multiple target location nodes and they are distributed in the target space grid, extract the maximum wind speed value and the average wind speed value from the wind speed forecast results of all target location nodes at the current prediction time. The maximum and average wind speed values ​​at multiple forecast times are continuously extracted to generate wind speed forecast time series curves. By comparing the wind speed forecast time series curve with the pre-stored safe extreme value range, when the maximum wind speed value is outside the safe extreme value range, wind speed warning information containing the spatial coordinate range and the expected occurrence time is generated.

9. A long short-term memory stochastic forest wind speed prediction system based on topographic influence, characterized in that, include: The parameter acquisition module is used to acquire historical wind speed sequences, environmental characteristic parameters, and terrain characteristic parameters for each wind measurement node, and perform normalization processing. The feature decoupling module is used to construct a time-series feature vector based on historical wind speed sequences, input the time-series feature vector into a long short-term memory network, obtain the prediction value of the long short-term memory network, and extract the short-term fluctuation feature components inside the long short-term memory network. The model mapping module is used to concatenate short-term fluctuation feature components, environmental feature parameters, and terrain feature parameters to obtain the model input vector, and then input the model input vector into the random forest model to obtain the random forest prediction value. The error allocation module is used to obtain the model bias based on the prediction values ​​of the long short-term memory network and the random forest, obtain the adaptive bias threshold based on the difference between adjacent data of historical wind speed sequences and terrain feature parameters, obtain the dynamic weight coefficient based on the adaptive bias threshold and the model bias, and generate the fusion prediction value of the wind measurement node based on the dynamic weight coefficient. The spatial reconstruction module is used to obtain the aerodynamic equivalent distance based on the vertical height difference and horizontal distance between the target location node and the wind measurement node, and to perform spatial interpolation processing on the fused prediction value based on the aerodynamic equivalent distance to obtain the wind speed forecast result of the target location node.

10. The long short-term memory random forest wind speed prediction system based on topographic influence according to claim 9, characterized in that, The error allocation module, when obtaining an adaptive deviation threshold based on the difference between adjacent data in the historical wind speed sequence and terrain feature parameters, is specifically used for: Extract the maximum absolute value of the difference between adjacent data in the historical wind speed sequence, and use the maximum value as the tolerance threshold base. Obtain the measured wind direction angle at the wind measurement node, as well as the normal azimuth and tilt angle in the terrain feature parameters; When the cosine of the difference between the measured wind direction angle and the normal azimuth angle is positive, the cosine value is multiplied by the sine of the tilt angle to obtain the penalty multiplier; when the cosine value is not positive, the penalty multiplier is set to zero. Add the value to the penalty factor to obtain the scaling factor, and multiply the tolerance threshold base by the scaling factor to obtain the adaptive deviation threshold.

Images