A wind power prediction method based on multi-source sensing data in extreme environment

Abstract

The application discloses a wind power prediction method based on multi-source sensing data in extreme environment, and comprises the following steps: step 1, determining the actual output of the wind turbine in the extreme environment according to the ideal active output of the wind turbine, distributing and fusing the weight based on the extreme working condition mutual information coupling degree, and realizing the collaborative processing of multi-source heterogeneous data through multi-source feature fusion; step 2, establishing a time-space feature extraction model for extreme working conditions, and mining the correlation features of multi-dimensional information; step 3, constructing a wind power prediction model to realize the accurate prediction of wind power; through the establishment of the multi-source feature fusion collaborative collection system, the application solves the problems of non-uniform time sequence, dimension conflict and difficult accurate elimination of extreme abnormality of multi-source heterogeneous data, can hierarchically dispose pulse drift and long-time disconnected defect data of various types, adaptively distributes and fuses the weight according to the coupling degree of features and output, fully releases the multi-dimensional monitoring information value, and makes up for the short board of information missing of a single data source.

Description

Technical Field

[0001] This invention belongs to the field of wind power prediction technology, specifically relating to a wind power prediction method based on multi-source sensor data under extreme environments. Background Technology

[0002] With the continued implementation of the dual-carbon strategy, wind power installed capacity has expanded rapidly, becoming a core clean power source in the new power system. Wind power output is constrained by multiple factors such as wind speed, temperature, icing, and turbine operating status, naturally exhibiting strong random fluctuations. Under extreme conditions such as cold waves, strong convection, extreme wind speeds, and blade icing, the spatiotemporal distribution of wind resources changes abruptly, and turbine mechanical losses and aerodynamic efficiency change synchronously. This significantly exacerbates the nonlinear and non-stationary characteristics of wind power output, leading to a substantial increase in power prediction errors. This directly increases the pressure on grid reserve configuration, threatening power system dispatch security and the efficiency of new energy consumption. Therefore, high-precision wind power forecasting under extreme environments has become an essential requirement for the industry.

[0003] Current wind power prediction technologies fall into two categories: traditional machine learning and deep learning. However, existing methods suffer from several inherent shortcomings: traditional time-series models, support vector machines, and shallow neural networks have limited feature mining capabilities, making it difficult to fit the strongly coupled nonlinear mapping relationships of multiple variables under extreme conditions; mainstream deep learning models such as LSTM and CNN mostly rely solely on a single data source, the wind turbine's SCADA system, neglecting multi-dimensional sensing information such as icing sensors, turbine vibration, and micro-meteorological data from the wind farm, thus failing to fully characterize the comprehensive impact mechanism of extreme weather on wind turbine output and exhibiting weak generalization ability in scenarios such as icing shutdowns and sudden gusts. Meanwhile, existing multi-source data fusion schemes have significant defects. Various sensor data vary greatly in scale, sampling time, and data quality. Conventional normalization and simple splicing methods cannot eliminate heterogeneous data conflicts, and there is a lack of targeted anomaly identification methods. Outliers caused by cold wave electromagnetic interference and low-temperature zero-point drift continuously interfere with the model training accuracy.

[0004] In addition, existing spatiotemporal feature mining methods are not adaptable enough, cannot dynamically adjust the spatial correlation weight of wind turbines according to wind direction, and are difficult to capture the power transfer pattern between stations under extreme wind conditions; conventional attention mechanisms do not provide differentiated gain for key disturbance features such as icing thickness and extreme wind speed, the training weight of extreme samples is too low, and they are easily dominated by stable operating condition samples, which ultimately leads to a significant increase in the prediction bias of cold wave and icing scenarios. Summary of the Invention

[0005] The purpose of this invention is to provide a wind power prediction method based on multi-source sensor data under extreme environments, so as to solve the problems mentioned in the background art.

[0006] To achieve the above objectives, the present invention provides the following technical solution: a wind power prediction method based on multi-source sensor data under extreme environments, comprising the following steps:

[0007] Step 1: Determine the actual output of the wind turbine in extreme environments based on the ideal active power output of the wind turbine, allocate fusion weights based on the mutual information coupling degree of extreme operating conditions, and achieve collaborative processing of multi-source heterogeneous data through multi-source feature fusion.

[0008] Step 2: Establish a spatiotemporal feature extraction model for extreme working conditions and mine the correlation features of multidimensional information;

[0009] Step 3: Construct a wind power prediction model to achieve accurate prediction of wind power.

[0010] Preferably, the method of allocating fusion weights based on the mutual information coupling degree under extreme operating conditions and achieving collaborative processing of multi-source heterogeneous data through multi-source feature fusion includes the following steps:

[0011] Step 11: Determine the ideal active power output of the wind turbine;

[0012] Step 12: Determine the actual output of the wind turbine in extreme environments;

[0013] Step 13: Construct a multi-source heterogeneous original dataset using multi-source features;

[0014] Step 13: Coarsely screen for drift anomalies through quartile intervals and then weight them. Detailed identification of specific anomalies;

[0015] Step 14: Assign fusion weights based on mutual information coupling degree under extreme working conditions, and determine the standardized feature matrix through multi-source feature fusion.

[0016] Preferably, the ideal active power output of the wind turbine is calculated as follows: ,in,

[0017] Contribute to the ideal of wind turbines;

[0018] Standard air density;

[0019] The sweep radius of the fan impeller;

[0020] The measured wind speed is based on the wheel hub height.

[0021] The wind energy utilization coefficient;

[0022] The tip speed ratio;

[0023] The blade pitch angle;

[0024] The actual output of the wind turbine in extreme environments is calculated as follows: ,in,

[0025] Actual power output of wind turbines in extreme environments;

[0026] Contribute to the ideal of wind turbines;

[0027] This is a correction factor for blade icing efficiency;

[0028] This is a correction factor for air density at high and low temperatures;

[0029] This refers to the vibration loss coefficient of the tower bearing.

[0030] Add power loss in extreme environments.

[0031] Preferably, the outlier value for coarse screening drift in the quartile interval is calculated as follows: ,in,

[0032] It is the lower quartile;

[0033] It is the upper quartile;

[0034] Interquartile range;

[0035] This is the upper coarse screening threshold;

[0036] The lower coarse screening threshold;

[0037] Weighted The calculation for precise identification of sensor drift anomalies is as follows: ,in,

[0038] This is the weighted adjusted sample standard deviation;

[0039] For extreme volatility weighting;

[0040] This represents the total number of sensor samples within a single sliding window.

[0041] This represents the k-th sensor sampling data point within the sliding window;

[0042] This is the average of all samples within the sliding window;

[0043] This represents the absolute deviation of a single sample from the moving window mean.

[0044] This is the critical threshold for anomaly detection.

[0045] Preferably, the calculation of the standardized feature matrix determined by multi-source feature fusion is as follows: , ,in,

[0046] This is the standardized feature matrix after multi-source feature fusion;

[0047] The fusion weighting coefficients for SCADA operation data;

[0048] The fusion weighting coefficient for meteorological monitoring data;

[0049] The fusion weighting coefficients for vibration monitoring data;

[0050] The fusion weighting coefficients for environmental perception data;

[0051] This is a subset of the SCADA operation data features of the wind turbine.

[0052] This is a subset of the meteorological monitoring data from the station.

[0053] This is a subset of the unit vibration monitoring data.

[0054] This is a subset of features from the leaf icing environment perception data.

[0055] For the normalized fusion weights of the i-th type of sensor data source;

[0056] Let be the set of sensing features of the i-th class.

[0057] Preferably, establishing a spatiotemporal feature extraction model for extreme working conditions and mining the correlation features of multidimensional information includes the following steps:

[0058] Step 21: Extract temporal features and wind field spatial correlation features from the standardized feature matrix after multi-source feature fusion. The temporal features and wind field spatial correlation features are spliced ​​together to form a high-dimensional hybrid feature.

[0059] Step 22: Use the maximum correlation and minimum redundancy criterion to select the optimal feature subset for modeling, balancing feature-output correlation and feature internal independence;

[0060] Step 23: Dynamically adjust the feature gain coefficient based on the working condition label to optimize the attention score allocation logic.

[0061] Preferably, the extracted time-series features are calculated as follows: ,in,

[0062] This is the time-series feature matrix;

[0063] It is a non-linear activation function;

[0064] This is the standardized feature matrix after multi-source feature fusion;

[0065] This is the convolution operator;

[0066] Temporal convolution kernel;

[0067] For bias terms;

[0068] The spatial correlation features of the wind field were extracted and calculated as follows: ,in,

[0069] This refers to the spatial correlation characteristics of wind fields;

[0070] It is a degree matrix;

[0071] This is a wind-weighted adjacency matrix;

[0072] These are spatially learnable weights.

[0073] Preferably, the optimal feature subset for modeling is selected using the maximum relevance and minimum redundancy criterion. ,in,

[0074] The objective function for the comprehensive evaluation of the maximum relevance and minimum redundancy criteria;

[0075] The objective function for the comprehensive evaluation of the maximum relevance and minimum redundancy criteria;

[0076] This is a subset of the overall features to be evaluated.

[0077] Actual power output of wind turbines in extreme environments;

[0078] Mutual information between the feature subset as a whole and wind power output;

[0079] This represents the total number of features contained within the current feature subset.

[0080] The i-th feature in the subset;

[0081] The j-th dimension feature in the subset;

[0082] The mutual information between the i-th and j-th dimensions;

[0083] The optimized attention score allocation logic is as follows: ,in,

[0084] Initial attention score for a single feature;

[0085] This represents the extreme gain coefficient.

[0086] Attention weights;

[0087] The transpose of the weight vector of the attention output layer;

[0088] For linear correction units;

[0089] The learnable weight matrix for the feature transformation layer;

[0090] Let i be the i-th spatiotemporal coupling feature vector of the input;

[0091] For learnable bias terms of the feature transformation layer;

[0092] This is for natural exponent calculations.

[0093] Preferably, constructing a wind power prediction model to achieve accurate wind power prediction includes the following steps:

[0094] Step 31: Select a bidirectional long short-term memory network to complete the weighted feature temporal decoding. Relying on forward and backward bidirectional gating units, it synchronously backtracks the historical output pattern and predicts the future output trend, adapting to the characteristics of sudden output changes and lag fluctuations under extreme working conditions.

[0095] Step 32: Construct an extreme weighted mean square error loss function, increase the loss weight of abnormal output samples, and improve the fitting accuracy of extreme output cliff and abrupt change samples.

[0096] Preferably, the gate control unit is calculated as follows: ,in,

[0097] In the forget gate state;

[0098] Input gate state;

[0099] The candidate cell state at time t;

[0100] This is a cellular memory state;

[0101] Use the Sigmoid activation function;

[0102] It is the hyperbolic tangent activation function;

[0103] ⊙ is for Hadamarda accumulation;

[0104] Here is the learnable weight matrix for the forget gate;

[0105] Let be the learnable weight matrix of the input gate;

[0106] The learnable weight matrix for the candidate cell state layer;

[0107] Learnable bias terms for forget gates;

[0108] For the learnable bias term of the input gate;

[0109] For learnable bias terms of candidate cell state layers;

[0110] Let be the hidden state vector of the previous time step t-1;

[0111] The spatiotemporal coupling features input at the current time t;

[0112] This represents the historical cell memory state at time t-1.

[0113] The extreme weighted mean square error loss function is calculated as follows: ,in,

[0114] This is the extreme weighted mean square error loss function;

[0115] This is the measured power.

[0116] To predict power;

[0117] These are extreme weighting coefficients.

[0118] The technical effects and advantages of this invention, a wind power prediction method based on multi-source sensor data under extreme environments:

[0119] 1. By building a multi-source feature fusion collaborative acquisition system and improving the quartile-3σ joint anomaly detection and mutual information weighted fusion preprocessing process, the system solves the problems of inconsistent time series, conflicting dimensions, and difficulty in accurately removing extreme anomalies in multi-source heterogeneous data. It can handle various defect data such as pulse drift and long-term disconnection in a layered manner, and adaptively allocate fusion weights according to the coupling degree between features and output, so as to fully release the value of multi-dimensional monitoring information and make up for the shortcomings of missing modeling information from a single data source.

[0120] 2. Through a dual-branch spatiotemporal feature extraction structure, which differs from traditional fixed topology networks, the spatial adjacency weights of wind turbines can be updated according to real-time wind direction, and the temporal mutation and spatial transmission patterns of the site can be mined simultaneously. Combined with the maximum correlation and minimum redundancy criteria, high-dimensional features are simplified, effectively eliminating redundant information, reducing the risk of overfitting, and significantly improving the ability to mine coupled features under extreme conditions.

[0121] 3. Construct an extreme adaptive multi-head attention and weighted loss prediction architecture, add an extreme working condition gain coefficient to amplify the weights of key features such as icing and strong winds, and adopt an extreme sample weighted loss function to balance the training ratio between stable and severe working conditions, thus overcoming the shortcomings of conventional models in fitting extreme scenarios. Attached Figure Description

[0122] Figure 1 This is a schematic diagram of the steps of the present invention;

[0123] Figure 2 This is a schematic diagram illustrating wind power prediction based on multi-source sensor data under extreme environments, as presented in this invention.

[0124] Figure 3 This is a schematic diagram of the collaborative preprocessing of multi-source heterogeneous sensor data according to the present invention;

[0125] Figure 4 This is a schematic diagram of the structure of the extreme adaptive attention wind power prediction model of the present invention;

[0126] Figure 5 This invention provides time series characteristics of multi-source sensor data for extreme weather processes.

[0127] Figure 6 This is a heatmap showing the correlation between the features of multi-source sensor data in this invention.

[0128] Figure 7 This is the convergence curve of the training and validation loss of the model in this invention;

[0129] Figure 8 This invention provides a comparison of wind power prediction results from different models under normal operating conditions.

[0130] Figure 9 This invention provides a comparison of wind power prediction results from different models under extreme operating conditions.

[0131] Figure 10 The probability density distribution of prediction errors for each model under different operating conditions in this invention;

[0132] Figure 11 This is a scatter plot of the model predictions under different working conditions of the present invention;

[0133] Figure 12 This is a comparison of evaluation metrics for different models of the present invention. Detailed Implementation

[0134] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0135] This invention provides, for example Figure 1 The wind power prediction method based on multi-source sensor data under extreme environments, as shown, includes the following steps:

[0136] Step 1: Construct a multi-source sensor data system that integrates SCADA operation data, meteorological monitoring data, vibration monitoring data, and environmental perception data. Determine the actual output of the wind turbine in extreme environments based on the ideal active power output of the wind turbine. Assign fusion weights based on the mutual information coupling degree under extreme operating conditions. Achieve collaborative processing of multi-source heterogeneous data through multi-source feature fusion.

[0137] Step 2: Establish a spatiotemporal feature extraction model for extreme operating conditions and mine the correlation features of multidimensional information (wind speed, wind direction, temperature, air pressure, unit vibration and operating status, etc.).

[0138] In this invention, the spatiotemporal feature extraction model for extreme operating conditions is specifically as follows:

[0139] 1. Input data settings and inherent relationships:

[0140] The input consists of standardized modeling data preprocessed from multiple sources of sensors. The data sources are divided into four categories: wind turbine operation monitoring, micro-meteorology at the wind farm, unit vibration, and blade icing. All data are synchronously matched with the same timestamp. There are significant heterogeneous differences within the data. Each type of data corresponds to four key factors affecting wind power output: wind turbine operating status, environmental disturbances, equipment losses, and blade icing. There is a strong nonlinear correlation between the data and the power generation to be predicted. Under four operating conditions—extreme cold waves, icing, gusts, and sudden changes in wind direction—the influence of environmental and icing sensor data on power output fluctuations is significantly enhanced, serving as the core input basis for distinguishing between stable and extreme operating conditions.

[0141] 2. Specific algorithm improvements and adaptation logic for wind power scenarios:

[0142] (1) To address the industry pain points of sudden power output changes under extreme wind power conditions, spatial wind turbine coupling, and multi-source data redundancy interference, a layered improvement was made. The overall structure is divided into four parts: time-series extraction branch, spatial extraction branch, extreme adaptive feature screening and attention gain module. The time-series extraction branch adopts a multi-scale one-dimensional convolution structure to adapt to extreme time-series disturbances at different time scales, such as short-term gust fluctuations, long-term cold wave cooling, and ice melting power cliffs, and automatically captures the power output change patterns at the time-series level.

[0143] (2) The spatial extraction branch adopts the convolutional structure of the wind direction dynamic topology graph, abandons the fixed wind turbine association topology, and dynamically adjusts the spatial association weight between wind turbines according to the real-time wind direction to match the spatial transfer characteristics of wind field power under extreme wind and sudden wind direction conditions.

[0144] (3) The feature selection module adopts the maximum relevance and minimum redundancy selection logic, giving priority to retaining features that are strongly related to wind power output and have low redundancy, eliminating invalid redundant dimensions, and adapting to extreme working conditions and small sample training scenarios.

[0145] (4) The extreme adaptive gating attention module uniformly configures the disturbance feature gain mechanism for four types of extreme working conditions, automatically amplifies the attention of extreme disturbance features such as temperature, icing thickness, gust rate, and wind direction change, weakens the regular operating features that dominate under stable working conditions, and solves the problem that extreme small sample features are covered by massive stable samples. Each module is deeply integrated with the wind power business logic. All improvements are designed around the fluctuation characteristics of wind power sensing data and the spatial coupling law of wind field under extreme weather conditions. There is no situation where general algorithms are not applied in a targeted manner.

[0146] 3. Output results:

[0147] The optimal spatiotemporal coupling feature set is output after filtering and optimization, redundant interference information is removed, and key disturbance information under extreme conditions is amplified simultaneously, serving as standardized input material for the backend wind power prediction model.

[0148] 4. Inherent relationship between input and output:

[0149] The input multi-source sensor raw data contains a large number of irrelevant, redundant, and stationary features. After the model performs wind power scenario-specific spatiotemporal mining and extreme adaptive weighting, the output features remove invalid dimensions and enhance the correlation information of extreme disturbances. The matching degree and correlation between the output features and the extreme power output fluctuations of wind turbines are much higher than those of the original input, which directly determines the recognition accuracy of the subsequent power prediction model for sudden power output changes under extreme operating conditions.

[0150] Step 3: Construct a wind power prediction model to achieve accurate prediction of wind power under complex operating conditions.

[0151] In this invention, the wind power prediction model is specifically as follows:

[0152] 1. Input data:

[0153] The input is the optimized spatiotemporal coupling features output by the aforementioned spatiotemporal feature extraction model. This feature set has completed the enhancement of extreme disturbance information and the removal of redundant features. It contains dual information of temporal fluctuation patterns and wind turbine spatial coupling, and is a standardized input material covering conventional and four types of extreme operating conditions.

[0154] 2. Model structure, complete training steps, and core parameter information:

[0155] (1) Complete model structure:

[0156] The main body adopts a bidirectional long short-term memory network to build the temporal decoding main body, and is combined with a front-end extreme adaptive attention weighting layer and a back-end fully connected fitting layer to form a complete prediction architecture. The bidirectional structure can simultaneously learn the historical temporal change trend and the future power output fluctuation pattern, and adapt to the lag fluctuation characteristics of power output sudden rise and fall under extreme conditions. The attention layer reuses the extreme condition gain logic of the spatiotemporal extraction module to further strengthen the weight of extreme disturbance features. The end fully connected layer realizes the mapping output of features to wind power values.

[0157] (2) Full-process training steps:

[0158] Step 1: Dataset partitioning. All time-series sensor samples are split into training, validation, and test sets in chronological order. This ensures that extreme condition samples are evenly distributed across subsets, preventing training bias caused by including all extreme samples in a single dataset. Step 2: Differentiated sample loss weighting. Higher loss weights are assigned to samples experiencing sudden output changes due to cold waves, icing, strong gusts, and sudden wind direction changes. Weights are reduced for samples with stable operation to prevent model training from being dominated by a large number of stable samples. Step 3: Iterative model training. The AdamW optimizer with a dynamically decaying learning rate is used for iterative training. After each iteration, the prediction error is checked on the validation set. If the validation set error does not decrease for several consecutive iterations, iteration is stopped to suppress overfitting to extreme small samples. Step 4: Model performance validation. The prediction error is tested separately under normal operating conditions and the four extreme conditions. The prediction deviation of traditional models in extreme scenarios is compared to validate the model's performance.

[0159] (3) Disclosure of key training parameters:

[0160] The system uniformly sets the upper limit of the number of training iterations, the initial learning rate, the learning rate decay rate, and the number of samples per batch; it distinguishes between normal and extreme operating conditions by setting two sets of loss weight ratios; it configures multiple sets of convolutional kernels of different sizes to cover long and short temporal fluctuations; it uniformly configures the feature gain coefficients for extreme operating conditions in the attention module, and fully records all business-related exclusive parameters.

[0161] 3. Output results:

[0162] It outputs accurate predicted power generation of wind turbines at the corresponding time scale. It can output low-error predicted values ​​under all operating conditions, including normal operation, low temperature cold wave, blade icing, extreme strong gusts, and sudden changes in wind direction. It can directly support the grid's new energy dispatch reserve capacity configuration.

[0163] 4. Technical effects and underlying principles:

[0164] In terms of technical performance, compared with traditional prediction schemes based on single-source data, fixed topology, and no extreme sample weighting, this model significantly reduces prediction errors under normal operating conditions and improves prediction accuracy even more under four extreme operating conditions. It can accurately capture sudden phenomena such as power output cliffs and short-term oscillations, and its generalization ability is adapted to extreme small sample scenarios. It greatly reduces the reserve capacity reserved by the power grid to cope with wind power output fluctuations and improves the level of new energy consumption.

[0165] The underlying principle is that a bidirectional long short-term memory network fully captures the temporal correlation, and two-stage extreme adaptive attention at the front end continuously amplifies the perturbation feature information. During the training phase, differentiated weighted loss forces the model to focus on learning the power output law under extreme conditions. Dynamic spatial topology matching matches the real spatial power transfer relationship of the wind field, thereby solving the core defects of existing technologies such as failure of extreme condition prediction and surge in error.

[0166] In this invention, the AdamW optimizer is used in conjunction with a dynamic learning rate decay strategy to suppress overfitting in extremely small sample sizes.

[0167] A wind farm in Northwest my country was selected as the data acquisition site. Four types of synchronous sensor acquisition units were deployed throughout the site. The sampling frequency was uniformly set to 1 minute / point, and the timestamps were aligned for synchronous acquisition. The sensor deployment parameters included wind turbine SCADA operation sensors, station micro meteorological sensors, vibration monitoring sensors, and environmental icing sensing sensors.

[0168] Key data acquisition features: real-time active power, reactive power, pitch angle, generator speed, nacelle tilt angle; hub wind speed, wind direction, ambient temperature, atmospheric pressure, air humidity; bearing triaxial vibration amplitude, tower low-frequency vibration frequency, vibration time-domain peak value; blade surface ice thickness, blade surface temperature, ice resistivity.

[0169] A total of 22 original input features were collected from four types of sensors. After removing zero-variance constant features, 18 effective modeling features were retained to construct a multi-source heterogeneous original dataset. The four types of sub-data sets have independent dimensions and significant differences in scale, requiring targeted heterogeneous preprocessing.

[0170] in:

[0171] This is a matrix of the original total dataset from multiple heterogeneous sources;

[0172] This is a subset of data for wind turbine SCADA operation monitoring.

[0173] This is a subset of data from a micro-meteorological sensor dataset for a field station.

[0174] This is a subset of the vibration sensing dataset;

[0175] This is a subset of the data set for sensing the icing state of the blades.

[0176] Specifically, the process of allocating fusion weights based on the mutual information coupling degree under extreme operating conditions and achieving collaborative processing of multi-source heterogeneous data through multi-source feature fusion includes the following steps:

[0177] Step 11: Determine the ideal active power output of the wind turbine;

[0178] Step 12: Determine the actual output of the wind turbine in extreme environments;

[0179] Step 13: Construct a multi-source heterogeneous original dataset using multi-source features;

[0180] Step 13: Coarsely screen for drift anomalies through quartile intervals and then weight them. Detailed identification of specific anomalies;

[0181] Step 14: Assign fusion weights based on mutual information coupling degree under extreme working conditions, and determine the standardized feature matrix through multi-source feature fusion.

[0182] In this invention, the 18-dimensional sensor features selected are paired one by one with the synchronously collected wind power output data. The frequency of occurrence of various types of samples is counted by dividing the continuous values ​​into intervals. By comparing the proportion of samples where the features and output change synchronously, the information contribution of each feature to the power generation is quantified to obtain the single feature coupling degree. Then, the coupling degrees of all features under the same type of sensor are accumulated to obtain the overall coupling degree of the four data sources: SCADA, meteorology, vibration, and icing.

[0183] The weight allocation adopts a two-level normalized weighting logic. The first layer first allocates the weight of the major categories, summing up the total coupling degree of the four types of sensors. The proportion of the total coupling degree of a single category to the total coupling degree of all coupling degrees is used as the overall weight of the data of that category. The major categories of sensors that are more closely related to the output will receive higher weights. The second layer allocates the fine weight of single features, summing up the total coupling degree of all 18-dimensional features. The proportion of the coupling degree of a single feature is used as an independent weighting coefficient, amplifying the proportion of highly correlated features such as icing thickness and extreme wind speed, and weakening low-correlation interference features.

[0184] During the fusion process, the entire group of sensor data is first scaled according to the major category weights, and then the values ​​of each column in the group are adjusted individually according to the single feature weights. After weighting, all data are spliced ​​together, and then uniform normalization is used to eliminate the difference in dimensions. Throughout the process, the weights are adaptively allocated based on the correlation between the data themselves. In extreme operating conditions, the weights of meteorological and icing data are automatically increased, while in normal operating conditions, the weights of unit operation data are emphasized to avoid key disturbance features being masked by massive amounts of stable data.

[0185] Specifically, based on the aerodynamic principle of wind energy conversion, under ideal operating conditions of standard atmospheric pressure, 20°C, and blades without icing deformation, the ideal active power output of a single wind turbine is calculated as follows: ,in,

[0186] The ideal power output of the wind turbine is kW;

[0187] For standard air density, take 1.225 kg / m³. ³;

[0188] The sweep radius of the fan impeller. ;

[0189] The measured wind speed is based on the wheel hub height. ;

[0190] The wind energy utilization coefficient;

[0191] The tip speed ratio;

[0192] The blade pitch angle;

[0193] In extreme environments, air density, blade aerodynamic shape, and unit transmission losses change synchronously. By introducing icing loss coefficients, temperature-density correction coefficients, and vibration transmission loss coefficients, and closely aligning with actual power output patterns, the actual power output of the wind turbine in extreme environments is calculated as follows: ,in,

[0194] Actual power output of the wind turbine in extreme environments, kW;

[0195] Contribute to the ideal of wind turbines;

[0196] This is a correction factor for blade icing efficiency; the larger the icing thickness, the smaller the value, with a range of 0.35-0.92.

[0197] This is a correction factor for air density at high and low temperatures. Under cold waves and low temperatures, the air density increases, so the factor is greater than 1. Under high temperatures and strong convection, the factor is less than 1.

[0198] The vibration loss coefficient of the tower bearing is the attenuation coefficient of the vibration exceeding the standard.

[0199] Additional power losses (yaw, pitch servo) are incurred in extreme environments (kW).

[0200] single The algorithm is only suitable for normally distributed sensor data, and the single quartile algorithm has weak ability to identify small drift anomalies. This application solves the problems of inconsistent time series, conflicting dimensions, and difficulty in accurately removing extreme anomalies in multi-source heterogeneous data by improving the quartile-3σ joint anomaly detection and mutual information weighted fusion preprocessing process. It can process various defective data such as pulse drift and long-term disconnection in a hierarchical manner, and adaptively allocate fusion weights according to the coupling degree of features and output, so as to fully release the value of multi-dimensional monitoring information and make up for the shortcomings of missing modeling information from a single data source.

[0201] Specifically, the outlier values ​​for the coarse screening in the quartile interval are calculated as follows: Calculate the quartiles of the one-dimensional sensing time series data, where

[0202] This represents the lower quartile of the dataset, corresponding to the 25th quartile x0.25, where x0.25 is the value at the 25th percentile of the sample sequence, outside the range. This is suspected to be a drift anomaly.

[0203] This represents the upper quartile of the dataset, corresponding to the 75th quartile x0.75, where x0.75 is the value at the 75th percentile of the sample sequence, outside the range. This is suspected to be a drift anomaly.

[0204] Interquartile range;

[0205] This is the upper coarse screening threshold;

[0206] The lower coarse screening threshold;

[0207] Weighted The calculation for precise identification of sensor drift anomalies is as follows: Introducing extreme volatility weighting coefficients Optimize the standard deviation threshold, where,

[0208] The weighted and corrected sample standard deviation is used to calculate the 3σ anomaly detection threshold;

[0209] For extreme fluctuation weighted system, extreme operating conditions Normal operating conditions This improves the sensitivity of identifying anomalies in extreme small fluctuations and meets the requirements for judging sensor drift anomalies.

[0210] This represents the total number of sensor samples within a single sliding window.

[0211] Let k be the k-th sensor sampling data point within the sliding window, where k is the index of the sample sequence. Traverse all window samples.

[0212] This is the average of all samples within the sliding window;

[0213] This represents the absolute deviation of a single sample from the moving window mean.

[0214] This is the critical threshold for anomaly detection.

[0215] (1) Single-point pulse anomaly: weighted interpolation completion of the nearest 5 points;

[0216] (2) Drift anomaly with a continuous duration of <10min: Time series moving average correction;

[0217] (3) Abnormal sensor disconnection with a continuous duration of ≥10 min: Based on multidimensional time-series KNN interpolation of adjacent sensors at the same location, data distortion under extreme working conditions is avoided.

[0218] Specifically, addressing the issue of conflicting dimensions among four types of heterogeneous sensors, the weights of low-coupling noise features are weakened, while the weights of highly sensitive features such as temperature, icing, and wind speed are strengthened. The standardized feature matrix is ​​calculated through multi-source feature fusion. , ,in,

[0219] The actual power output of the wind turbine in extreme environments, in kW, is used as the prediction target label for mutual information calculation.

[0220] This is the standardized feature matrix after multi-source feature fusion;

[0221] The fusion weighting coefficients for SCADA operation data;

[0222] The fusion weighting coefficient for meteorological monitoring data;

[0223] The fusion weighting coefficients for vibration monitoring data;

[0224] The fusion weighting coefficients for environmental perception data;

[0225] This is a subset of the preprocessed and normalized SCADA operating data of the wind turbine.

[0226] This is a subset of the features of the preprocessed and normalized meteorological monitoring data from the stations;

[0227] This is a feature subset of the unit vibration monitoring data after preprocessing and normalization;

[0228] This is a subset of features from the preprocessed and normalized leaf icing environment perception data;

[0229] For the normalized fusion weights of the i-th type of sensor data source;

[0230] Let be the set of sensing features of the i-th class.

[0231] The features are after Min-Max normalization. The normalization formula is calculated as follows: ,in,

[0232] These are the standardized feature values ​​after Min-Max normalization.

[0233] These are the original, unnormalized single sensor feature sample values;

[0234] It is the minimum value among all samples of this feature column;

[0235] This is the maximum value among all samples of this feature column.

[0236] Global feature mapping to the [0,1] interval eliminates the differences in physical dimensions of voltage, vibration thickness, and temperature, thereby improving the convergence speed of subsequent deep learning.

[0237] Specifically, establishing a spatiotemporal feature extraction model for extreme working conditions and mining the correlation features of multidimensional information includes the following steps:

[0238] Step 21: Extract temporal features and wind field spatial correlation features from the standardized feature matrix after multi-source feature fusion. The temporal features and wind field spatial correlation features are spliced ​​together to form a high-dimensional hybrid feature.

[0239] Step 22: Use the maximum correlation and minimum redundancy criterion to select the optimal feature subset for modeling, balancing feature-output correlation and feature internal independence;

[0240] Step 23: Dynamically adjust the feature gain coefficient based on the working condition label to optimize the attention score allocation logic.

[0241] Specifically, 1D-CNN is used to mine multi-dimensional coupled features of historical time series data from a single camera position. For the pre-processed 18-dimensional standardized fused time series features, a three-layer convolutional pooling stacked structure is constructed to specifically extract two types of time series features under extreme conditions: short-term abrupt changes and long-term trends. This is adapted to the characteristics of cold wave cooling, ice melting power output cliffs, and short-term fluctuations in gusts. The extracted time series features are calculated as follows: ,in,

[0242] This is the time-series feature matrix;

[0243] It is a nonlinear activation function that fits strongly nonlinear mapping relationships in extreme environments;

[0244] This is the standardized feature matrix after multi-source feature fusion;

[0245] This is the convolution operator;

[0246] The kernel is a temporal convolution kernel with a size set to [3,5,7] to accommodate extreme fluctuations at different periods.

[0247] For bias terms;

[0248] Based on the measured geographic coordinates of wind turbine locations, an undirected spatial topological graph G(V,E) is constructed, where V is the set of wind turbine nodes and E is the set of wind-direction connected edges. The adjacency weight matrix is ​​dynamically updated based on real-time wind direction, conforming to the spatial transfer patterns of wind resources in severe convective weather. The spatial correlation features of the wind field are extracted and calculated using GCN graph convolution. ,in,

[0249] This refers to the spatial correlation characteristics of wind fields;

[0250] It is a degree matrix;

[0251] This is a wind-weighted adjacency matrix;

[0252] With spatially learnable weights, dynamic adjacency matrices can significantly improve the accuracy of spatial feature extraction under extreme wind direction conditions compared to fixed topological graph convolution.

[0253] The spatiotemporal bi-branch features, when concatenated, form high-dimensional hybrid features, which suffer from dimensional redundancy and feature autocorrelation interference, easily leading to overfitting and computational redundancy. To address this, the maximum correlation and minimum redundancy criterion is used to select the optimal feature subset for modeling, balancing feature-output correlation with feature internal independence. Specifically, the calculation for selecting the optimal feature subset using the maximum correlation and minimum redundancy criterion is as follows: ,in,

[0254] The objective function for the comprehensive evaluation of the maximum relevance and minimum redundancy criteria is denoted by , where a larger value indicates a better quality feature subset.

[0255] The comprehensive evaluation objective function is the maximum relevance and minimum redundancy criterion. The maximum value of the evaluation function is obtained to complete the selection of the optimal feature subset.

[0256] This is a subset of the overall features to be evaluated.

[0257] The actual output of wind turbines in extreme environments is the target variable for prediction;

[0258] The mutual information between the feature subset as a whole and the wind power output represents the amount of information related to the output by the feature;

[0259] This represents the total number of features contained within the current feature subset.

[0260] The i-th feature in the subset;

[0261] The j-th dimension feature in the subset;

[0262] The mutual information between the i-th and j-th dimensions measures the redundancy between the two sets of features.

[0263] In the formula: the first term maximizes the correlation between the feature and the measured wind power output, and the second term minimizes the internal redundancy of the feature. After iterative screening, the 12-dimensional optimal spatiotemporal coupling feature is retained and input into the back-end prediction network, which effectively reduces the training computing power and is suitable for training under extreme small sample conditions.

[0264] The optimized attention score allocation logic is as follows: ,in,

[0265] Initial attention score for a single feature;

[0266] For extreme gain coefficients in normal operating conditions Four types of extreme working conditions are uniformly set ;

[0267] The four extreme operating conditions are:

[0268] Low temperature cold wave conditions: The ambient temperature drops sharply and the air density changes, causing the unit output baseline to shift.

[0269] Blade icing condition: Blade icing causes aerodynamic losses and a sharp drop in power output;

[0270] Extreme strong gust conditions: short-term wind speed fluctuates drastically, and frequent operation of the pitch system causes power output oscillations.

[0271] Severe convective thunderstorms and strong winds: rapid changes in wind direction, a surge in turbulence intensity, and disordered power transmission in the wind field.

[0272] To normalize attention weights and achieve adaptive weighting of features sensitive to harsh environments;

[0273] The transpose of the weight vector of the attention output layer;

[0274] For linear correction units, it is an activation function that introduces a nonlinear transformation;

[0275] The learnable weight matrix for the feature transformation layer;

[0276] Let i be the i-th spatiotemporal coupling feature vector of the input;

[0277] For learnable bias terms of the feature transformation layer;

[0278] This is the natural exponent operation, used for Softmax normalization mapping.

[0279] Specifically, constructing a wind power prediction model to achieve accurate wind power prediction includes the following steps:

[0280] Step 31: Select a bidirectional long short-term memory network to complete the weighted feature temporal decoding. Relying on forward and backward bidirectional gating units, it synchronously backtracks the historical output pattern and predicts the future output trend, adapting to the characteristics of sudden output changes and lag fluctuations under extreme working conditions.

[0281] Step 32: Construct an extreme weighted mean square error loss function, increase the loss weight of abnormal output samples, and improve the fitting accuracy of extreme output cliff and abrupt change samples.

[0282] Specifically, the calculation for a single gate unit is as follows: ,in,

[0283] In the forget gate state;

[0284] Input gate state;

[0285] The candidate cell state at time t;

[0286] This is a cellular memory state;

[0287] Use the Sigmoid activation function;

[0288] It is the hyperbolic tangent activation function;

[0289] ⊙ is for Hadamarda accumulation;

[0290] Here is the learnable weight matrix for the forget gate;

[0291] Let be the learnable weight matrix of the input gate;

[0292] The learnable weight matrix for the candidate cell state layer;

[0293] Learnable bias terms for forget gates;

[0294] For the learnable bias term of the input gate;

[0295] For learnable bias terms of candidate cell state layers;

[0296] Let be the hidden state vector of the previous time step t-1;

[0297] The spatiotemporal coupling features input at the current time t;

[0298] This represents the historical cell memory state at time t-1.

[0299] The extreme weighted mean square error loss function is calculated as follows: ,in,

[0300] This is the extreme weighted mean square error loss function;

[0301] This is the measured power.

[0302] To predict power;

[0303] These are extreme weighting coefficients.

[0304] Example: This example demonstrates the implementation of this method at a 100MW wind farm in Northwest my country. Four types of sensors were deployed simultaneously: wind turbine SCADA, micro-meteorology, main shaft vibration, and blade icing sensors. A unified 1-minute sampling frequency was used, and timestamp alignment was completed to collect multi-dimensional raw data including power, wind speed, temperature, vibration, and icing thickness. An improved quartile-3σ joint algorithm was employed to identify three types of anomalies: pulse, drift, and prolonged disconnection. Data completion was performed using weighted interpolation, moving average, and multi-dimensional KNN based on the fault duration. Finally, weights were allocated based on mutual information to fuse multi-source features, and Min-Max normalization was applied. A spatiotemporal dual-branch feature extraction network was constructed, and a 1D-CNN was used to extract temporal fluctuation features. A dynamic wind direction-weighted GCN was used to mine the spatial correlation of wind turbines, and the optimal feature subset was selected based on the maximum correlation and minimum redundancy criterion. An extreme adaptive multi-head attention mechanism was introduced to strengthen the feature weights under severe operating conditions. An extreme weighted loss BiLSTM network was constructed to complete power prediction. The AdamW optimizer was used for iterative training, and the final output was the advanced wind power prediction result. The model performance was verified by RMSE, MAE, and R².

[0305] To verify the effectiveness of the method of the present invention (MS-STLSTM), a 100MW wind farm in Northwest my country was selected as the research object. The wind farm has an installed capacity of 100MW and a total of 50 horizontal axis wind turbine generators with a single unit capacity of 2.0 MW. The wind turbine hub height is 80 m, the rotor diameter is 96 m, the cut-in wind speed is 3 m / s, the rated wind speed is 12 m / s, and the cut-out wind speed is 25 m / s.

[0306] The data collection period was from January 2023 to March 2024, with a sampling interval of 15 minutes, and a total of 52,608 valid data samples were collected.

[0307] The multi-source sensor data system adopted includes the following four types of data: (1) SCADA operation data, including active power, reactive power, generator speed, blade pitch angle, nacelle temperature, gearbox temperature and other unit operation status parameters; (2) meteorological monitoring data, including meteorological elements such as wind speed, wind direction, ambient temperature, atmospheric pressure, relative humidity and other meteorological elements; (3) vibration monitoring data, including nacelle front-to-back vibration acceleration, left-to-right vibration acceleration and gearbox vibration velocity and other vibration parameters; (4) environmental perception data, including icing status indicators, dust concentration, lightning activity level and other environmental characteristic parameters.

[0308] The raw multi-source sensor data contains problems such as missing values, outliers, and inconsistent dimensions, which require systematic preprocessing.

[0309] Figure 5 This paper presents the time-series characteristics of multi-source sensor data during a typical extreme weather event (December 15-18, 2023). Figure 5 (a) It can be seen that the wind speed fluctuated drastically during the 30-40h period, with the maximum wind speed reaching 21.3m / s, which exceeded the wind speed threshold for the wind turbine to cut off, causing the wind turbine to enter the protection shutdown state. Figure 5 (b) It shows that the wind power output drops sharply to near zero during periods of extreme wind speed, exhibiting a significant nonlinear correlation with wind speed changes; Figure 5 (c) indicates that the ambient temperature dropped sharply by 12.6℃ within 30-45 hours, forming a typical cold wave weather process; Figure 5 (d) shows that atmospheric pressure exhibits a significant downward trend during extreme weather events, reaching a minimum of 992.3 hPa; Figure 5 (e) The vibration amplitude of the unit increases significantly under extreme wind speed conditions, with the maximum vibration velocity reaching 16.2 mm / s, approaching the alarm threshold. The complex coupling relationship between the above multi-source data fully illustrates the enormous challenges faced in wind power prediction under extreme environments.

[0310] To fully explore the potential correlation characteristics among multi-source sensor data, Pearson correlation coefficient matrix was used to conduct correlation analysis on each feature variable. The results are as follows: Figure 6 As shown, by Figure 6It can be seen that: (1) wind speed is strongly positively correlated with active power (r=0.92), which is the most critical factor affecting wind power output; (2) the X / Y direction of unit vibration is significantly positively correlated with wind speed (r=0.72 and 0.68 respectively), which can effectively reflect the mechanical load state of the unit under extreme wind conditions; (3) ambient temperature is strongly positively correlated with nacelle temperature and gearbox temperature (r=0.65 and 0.48 respectively), which can be used to characterize the changes in the thermal state of the unit; (4) atmospheric pressure is moderately negatively correlated with wind speed (r=-0.45), which has a certain early warning effect on extreme weather processes; (5) blade pitch angle is moderately negatively correlated with wind speed (r=-0.55), which reflects the active control characteristics of the pitch system in the wind speed regulation process. The above feature correlation analysis results provide an important basis for the subsequent selection of model input features.

[0311] The network structure parameters of the MS-STLSTM model in this invention are set as follows: The spatiotemporal feature extraction module contains a 3-layer spatiotemporal graph convolutional network (ST-GCN), with each layer containing 64, 128, and 256 graph convolutional kernels, and the graph adjacency matrix is ​​constructed using an adaptive learning method based on feature similarity; the temporal dependency modeling module uses a 2-layer bidirectional LSTM network with 256 and 128 hidden layer units, respectively; the multi-source feature fusion module adopts an attention-based fusion strategy with 8 attention heads; the prediction output module contains a 2-layer fully connected network with 128 and 1 neurons, respectively. The input sequence length is 96 (corresponding to 24 hours of historical data), and the prediction stride is 4 (corresponding to 1 hour of advance prediction).

[0312] To verify the performance gain of multi-source data fusion on the model, the following four comparison schemes were set up: (1) MS-STLSTM model of the present invention, which fuses all four types of multi-source sensor data; (2) traditional LSTM model, which uses only SCADA running data as input; (3) BP neural network model, with input features the same as scheme (1); (4) Support vector machine regression (SVM) model, with input features the same as scheme (1). The network structure and hyperparameters of each comparison model were optimized and selected by grid search method.

[0313] Figure 7 The loss convergence curves of each model during training are shown. Figure 7 (a) It can be seen that the MS-STLSTM model exhibits the fastest convergence speed and the lowest steady-state loss during training. After approximately 60 training rounds, the training loss tends to stabilize, eventually converging to around 0.002. The traditional LSTM model has a relatively slower convergence speed, stabilizing after approximately 100 training rounds, with a steady-state loss of approximately 0.010. The BP neural network model has the slowest convergence speed and exhibits significant oscillations, with the loss value still fluctuating around 0.025 after approximately 150 training rounds. Figure 7As can be seen from the validation loss curve in (b), the MS-STLSTM model also exhibits the best generalization performance on the validation set, without any obvious overfitting, while the BP neural network model shows a certain degree of overfitting in the later stages of training.

[0314] To comprehensively evaluate the predictive performance of each model, two typical scenarios—normal and extreme operating conditions—were selected for testing and analysis. The normal operating condition test set consisted of data from a period of stable wind speeds in June 2023, totaling 96 test samples; the extreme operating condition test set consisted of data from a period of cold waves accompanied by strong winds in December 2023, totaling 144 test samples.

[0315] Figure 8 The comparison curves of wind power prediction results for each model under normal operating conditions are shown. Figure 8 It can be seen that the MS-STLSTM model proposed in this paper best matches the actual power curve, accurately capturing the intraday periodic variation characteristics of wind power. The traditional LSTM model has a slightly better prediction performance, but exhibits a significant lag bias during periods of drastic power fluctuations. The BP neural network model and SVM regression model have large prediction biases, especially during periods of rapid power changes, making it difficult to accurately track the actual power trend. These results demonstrate that, under normal operating conditions, the combination of multi-source data fusion and deep learning networks can effectively improve the accuracy of wind power prediction.

[0316] Figure 9 The comparison curves of wind power prediction results for various models under extreme operating conditions are presented. Figure 9 It can be seen that during the 12-16 hour period, due to extreme wind speeds exceeding the cut-out threshold, the wind turbines entered a protective shutdown state, and the power output dropped sharply to near zero. Under these extreme conditions, the MS-STLSTM model of this invention can still predict the sudden power change process relatively accurately. This is mainly due to the model's full integration of multi-source sensor information such as unit vibration and atmospheric pressure, allowing it to anticipate the occurrence trend of extreme weather. In contrast, the traditional LSTM model shows a significant increase in prediction bias during the wind turbine cut-out and recovery phases, while the prediction results of the BP neural network and SVM models show serious deviations and cannot effectively adapt to the nonlinear changing characteristics under extreme operating conditions. During the recovery phase from 16-22 hours, the MS-STLSTM model also demonstrates strong adaptability and can quickly track the power recovery process.

[0317] Figure 10 The probability density distribution of prediction errors for each model is presented under both normal and extreme operating conditions. Figure 10(a) It can be seen that under normal operating conditions, the prediction error distribution of the MS-STLSTM model is the most concentrated, with the error peak located near zero, exhibiting an approximately symmetrical normal distribution, indicating that the model has good unbiasedness; the error distribution of the LSTM model is relatively dispersed; while the error distributions of the BP neural network and SVM models show obvious wide-tailed characteristics, indicating that these two models have significant prediction biases. Figure 10 (b) It can be seen that under extreme conditions, the error distribution of each model has broadened, but the error distribution of the MS-STLSTM model remains relatively concentrated, and the error standard deviation is significantly smaller than that of other comparative models; while the error distribution of the BP neural network and SVM models shows serious skewness under extreme conditions, indicating that these two models have weak adaptability to extreme environments.

[0318] Figure 11 A scatter plot showing the predicted and actual power values ​​is presented. Figure 11 (a) It can be seen that under normal operating conditions, the prediction scatter points of the MS-STLSTM model are closely distributed near the ideal prediction line (y=x), indicating that the predicted values ​​are highly consistent with the actual values; Figure 11 (b) It can be seen that, under extreme conditions, although the prediction scatter points of the MS-STLSTM model show a certain degree of dispersion, they are still well distributed along the ideal prediction line. No obvious systematic deviations are observed when the power is close to zero (wind turbine cut-out) and near the rated power, which further verifies the good prediction performance of the proposed method under extreme conditions.

[0319] To further quantitatively evaluate the predictive performance of each model, the root mean square error (RMSE), mean absolute error (MAE), and coefficient of determination (R²) were used as evaluation indicators. The comparison results of the evaluation indicators of each model under normal and extreme operating conditions are shown in Table 1.

[0320] Table 1 Comparison of Evaluation Indicators for Wind Power Prediction of Different Models

[0321] MS-STLSTM 0.065 0.052 0.972 0.098 0.078 0.938 LSTM 0.128 0.098 0.892 0.195 0.152 0.758 BP neural network 0.215 0.168 0.698 0.328 0.258 0.318 SVM 0.268 0.212 0.528 0.425 0.338 0.148

[0322] As shown in Table 1, under normal operating conditions, the MS-STLSTM model of this invention has an RMSE of 0.065MW, a MAE of 0.052MW, and an R² of 0.972, all of which are superior to the comparative models. Compared with the LSTM model using a single SCADA data input, the MS-STLSTM model reduces RMSE by 49.2%, MAE by 46.9%, and R² by 9.0%; compared with the BP neural network model using the same multi-source data input, RMSE reduces by 69.8%, MAE by 69.0%, and R² by 39.3%; compared with the SVM regression model, RMSE reduces by 75.7%, MAE by 75.5%, and R² by 84.1%. These results demonstrate that the fusion of multi-source sensor data can effectively improve the accuracy of wind power prediction, and the spatiotemporal feature extraction and deep learning methods proposed in this paper have stronger feature learning capabilities than traditional machine learning algorithms.

[0323] Under extreme conditions, the prediction accuracy of all models decreased, but the MS-STLSTM model maintained the best prediction performance, with an RMSE of 0.098 MW, a MAE of 0.078 MW, and an R² of 0.938. Compared with the LSTM model, the RMSE decreased by 49.7%, the MAE decreased by 48.7%, and the R² increased by 23.7%; compared with the BP neural network model, the RMSE decreased by 70.1%, the MAE decreased by 69.8%, and the R² increased by 194.9%; compared with the SVM model, the RMSE decreased by 76.9%, the MAE decreased by 76.9%, and the R² increased by 533.1%. It is particularly noteworthy that under extreme conditions, the R² of the BP neural network and SVM models dropped to 0.318 and 0.148, respectively, almost losing their predictive ability, while the MS-STLSTM model, thanks to its advantages in multi-source data fusion and spatiotemporal feature extraction, maintained high prediction accuracy, fully validating the strong adaptability of the proposed method to extreme environments.

[0324] Figure 12 The comparison results of the three evaluation indicators for each model under normal and extreme operating conditions are presented intuitively. Figure 12 (a) shows that the MS-STLSTM model has significantly lower RMSE than other models under both normal and extreme operating conditions, and exhibits the least performance degradation due to changes in operating conditions; Figure 12 (b) It can be seen that the MAE indicator exhibits a similar pattern to RMSE; from Figure 12 (c) It can be seen that the R² of the MS-STLSTM model remains above 0.93 under both normal and extreme conditions, demonstrating excellent predictive stability and robustness.

[0325] To further verify the contribution of multi-source sensor data fusion to prediction performance, ablation experiments were conducted on the MS-STLSTM model to test the prediction performance under four input configurations: using only SCADA data (Scheme A), SCADA + meteorological data (Scheme B), SCADA + meteorological + vibration data (Scheme C), and all four types of data (Scheme D, i.e., the complete model proposed in this paper).

[0326] Table 2 shows the comparison of prediction performance of different data configuration schemes under normal and extreme operating conditions. As can be seen from Table 2: (1) Under normal operating conditions, compared with using only SCADA data (Scheme A), gradually increasing meteorological data, vibration data and environmental perception data reduced RMSE by 18.0%, 12.3% and 8.2% respectively, indicating that all types of data sources have a positive contribution to prediction accuracy; (2) Under extreme operating conditions, the gain effect of multi-source data fusion is more significant. Compared with Scheme A, the RMSE of Scheme D is reduced by 41.3%, indicating that in extreme environments, multi-source data fusion plays a key role in improving the model's adaptability; (3) Meteorological data and vibration data contribute the most to improving prediction performance under extreme operating conditions, which is consistent with the physical mechanism of drastic changes in meteorological parameters and abnormal aggravation of unit vibration during extreme weather.

[0327] Table 2 Comparison of Prediction Performance for Different Data Configuration Schemes

[0328] Option A SCADA data only 0.111 0.167 0.845 Option B SCADA+ meteorological data 0.091 0.132 0.902 Option C SCADA + Weather + Vibration 0.080 0.112 0.925 Option D All four types of data 0.065 0.098 0.938

[0329] In summary, the wind power prediction method based on multi-source sensor data proposed in this invention for extreme environments constructs a spatiotemporal feature extraction model for extreme environments by integrating SCADA operation data, meteorological monitoring data, vibration monitoring data, and environmental perception data. This effectively uncovers deep correlations between heterogeneous multi-source data, significantly improving the prediction accuracy and model robustness of wind power under complex operating conditions. Experimental results show that the proposed method outperforms traditional machine learning models and single-source data prediction models under both normal and extreme operating conditions, providing reliable technical support for the safe and stable operation of high-proportion renewable energy power systems.

[0330] Actual measurements show that, regardless of whether it is a normal weather condition or a cold wave, icing, or extreme wind speed conditions, the RMSE and MAE of this invention are much lower than those of traditional models such as LSTM, BP, and SVM. The coefficient of determination is significantly improved, and the model's generalization ability and robustness against interference are comprehensively superior to existing prediction schemes.

[0331] Finally, it should be noted that the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A wind power prediction method based on multi-source sensor data under extreme environments, characterized in that, Includes the following steps: Step 1: Determine the actual output of the wind turbine in extreme environments based on the ideal active power output of the wind turbine, allocate fusion weights based on the mutual information coupling degree of extreme operating conditions, and achieve collaborative processing of multi-source heterogeneous data through multi-source feature fusion. Step 2: Establish a spatiotemporal feature extraction model for extreme working conditions and mine the correlation features of multidimensional information; Step 3: Construct a wind power prediction model to achieve accurate prediction of wind power.

2. The wind power prediction method based on multi-source sensor data under extreme environments according to claim 1, characterized in that, Based on the mutual information coupling degree under extreme operating conditions, the fusion weights are assigned, and multi-source heterogeneous data collaborative processing is achieved through multi-source feature fusion, including the following steps: Step 11: Determine the ideal active power output of the wind turbine; Step 12: Determine the actual output of the wind turbine in extreme environments; Step 13: Construct a multi-source heterogeneous original dataset using multi-source features; Step 13: Coarsely screen for drift anomalies through quartile intervals and then weight them. Detailed identification of specific anomalies; Step 14: Assign fusion weights based on mutual information coupling degree under extreme working conditions, and determine the standardized feature matrix through multi-source feature fusion.

3. The wind power prediction method based on multi-source sensor data under extreme environments according to claim 2, characterized in that: The ideal active power output of the wind turbine is calculated as follows: ,in, Contribute to the ideal of wind turbines; Standard air density; The sweep radius of the fan impeller; The measured wind speed is based on the wheel hub height. The wind energy utilization coefficient; The tip speed ratio; The blade pitch angle; The actual output of the wind turbine in extreme environments is calculated as follows: ,in, Actual power output of wind turbines in extreme environments; Contribute to the ideal of wind turbines; This is a correction factor for blade icing efficiency; This is a correction factor for air density at high and low temperatures; This refers to the vibration loss coefficient of the tower bearing. Add power loss in extreme environments.

4. The wind power prediction method based on multi-source sensor data under extreme environments according to claim 2, characterized in that: The calculation of coarse screening drift anomalies in the interquartile range is as follows: ,in, It is the lower quartile; It is the upper quartile; Interquartile range; This is the upper coarse screening threshold; The lower coarse screening threshold; Weighted The calculation for precise identification of sensor drift anomalies is as follows: ,in, This is the weighted adjusted sample standard deviation; For extreme volatility weighting; This represents the total number of sensor samples within a single sliding window. This represents the k-th sensor sampling data point within the sliding window; This is the average of all samples within the sliding window; This represents the absolute deviation of a single sample from the moving window mean. This is the critical threshold for anomaly detection.

5. The wind power prediction method based on multi-source sensor data under extreme environments according to claim 2, characterized in that: The standardized feature matrix is ​​calculated by multi-source feature fusion. , ,in, This is the standardized feature matrix after multi-source feature fusion; The fusion weighting coefficients for SCADA operation data; The fusion weighting coefficient for meteorological monitoring data; The fusion weighting coefficients for vibration monitoring data; The fusion weighting coefficients for environmental perception data; This is a subset of the SCADA operation data features of the wind turbine. This is a subset of the meteorological monitoring data from the station. This is a subset of the unit vibration monitoring data. This is a subset of features from the leaf icing environment perception data. For the normalized fusion weights of the i-th type of sensor data source; Let be the set of sensing features of the i-th class.

6. The wind power prediction method based on multi-source sensor data under extreme environments according to claim 1, characterized in that, Establishing a spatiotemporal feature extraction model for extreme working conditions and mining the correlation features of multidimensional information includes the following steps: Step 21: Extract temporal features and wind field spatial correlation features from the standardized feature matrix after multi-source feature fusion. The temporal features and wind field spatial correlation features are spliced ​​together to form a high-dimensional hybrid feature. Step 22: Use the maximum correlation and minimum redundancy criterion to select the optimal feature subset for modeling, balancing feature-output correlation and feature internal independence; Step 23: Dynamically adjust the feature gain coefficient based on the working condition label to optimize the attention score allocation logic.

7. The wind power prediction method based on multi-source sensor data under extreme environments according to claim 6, characterized in that: The extracted time-series features are calculated as follows: ,in, This is the time-series feature matrix; It is a non-linear activation function; This is the standardized feature matrix after multi-source feature fusion; This is the convolution operator; Temporal convolution kernel; For bias terms; The spatial correlation features of the wind field were extracted and calculated as follows: ,in, This refers to the spatial correlation characteristics of wind fields; It is a degree matrix; This is a wind-weighted adjacency matrix; These are spatially learnable weights.

8. The wind power prediction method based on multi-source sensor data under extreme environments according to claim 6, characterized in that: The optimal feature subset for modeling is selected using the maximum relevance and minimum redundancy criterion. ,in, The objective function for the comprehensive evaluation of the maximum relevance and minimum redundancy criteria; The objective function for the comprehensive evaluation of the maximum relevance and minimum redundancy criteria; This is a subset of the overall features to be evaluated. Actual power output of wind turbines in extreme environments; Mutual information between the feature subset as a whole and wind power output; This represents the total number of features contained within the current feature subset. The i-th feature in the subset; The j-th dimension feature in the subset; The mutual information between the i-th and j-th dimensions; The optimized attention score allocation logic is as follows: ,in, Initial attention score for a single feature; This represents the extreme gain coefficient. Attention weights; The transpose of the weight vector of the attention output layer; For linear correction units; The learnable weight matrix for the feature transformation layer; Let i be the i-th spatiotemporal coupling feature vector of the input; For learnable bias terms of the feature transformation layer; This is for natural exponent calculations.

9. The wind power prediction method based on multi-source sensor data under extreme environments according to claim 1, characterized in that, The steps involved in constructing a wind power prediction model to achieve accurate wind power prediction are as follows: Step 31: Select a bidirectional long short-term memory network to complete the weighted feature temporal decoding. Relying on forward and backward bidirectional gating units, it synchronously backtracks the historical output pattern and predicts the future output trend, adapting to the characteristics of sudden output changes and lag fluctuations under extreme working conditions. Step 32: Construct an extreme weighted mean square error loss function, increase the loss weight of abnormal output samples, and improve the fitting accuracy of extreme output cliff and abrupt change samples.

10. A wind power prediction method based on multi-source sensor data under extreme environments according to claim 9, characterized in that: The gating unit is calculated as follows: ,in, In the forget gate state; Input gate state; The candidate cell state at time t; This is a cellular memory state; Use the Sigmoid activation function; It is the hyperbolic tangent activation function; ⊙ is for Hadamarda accumulation; Here is the learnable weight matrix for the forget gate; Let be the learnable weight matrix of the input gate; The learnable weight matrix for the candidate cell state layer; Learnable bias terms for forget gates; For the learnable bias term of the input gate; For learnable bias terms of candidate cell state layers; Let be the hidden state vector of the previous time step t-1; The spatiotemporal coupling features input at the current time t; This represents the historical cell memory state at time t-1. The extreme weighted mean square error loss function is calculated as follows: ,in, This is the extreme weighted mean square error loss function; This is the measured power. To predict power; These are extreme weighting coefficients.

Images