An ultra-short-term wind power prediction method, device and system

Abstract

The application relates to the technical field of wind power prediction, in particular to an ultra-short-term wind power prediction method, device and system, which comprises the following steps: acquiring wind speed data and active power of each wind turbine; dividing the same wind condition cluster based on spatial distribution and historical wind speed, combining power difference at adjacent time to construct a local jump index, and constructing an abnormal fluctuation evaluation index according to the deviation of the index in the cluster; determining an evaluation coefficient based on the fluctuation difference of wind speed and active power and the abnormal fluctuation evaluation index; determining a training loss weight according to the evaluation coefficient and training a prediction model to realize wind power prediction. The application solves the problem that abnormal fluctuation caused by abnormal interference in wind power historical data and wind speed-power coupling mismatch make the model learn pseudo rules and reduce prediction accuracy, and improves the accuracy of ultra-short-term wind power prediction.

Description

Technical Field

[0001] This application relates to the field of wind power prediction technology, specifically to a method, device and system for ultra-short-term wind power prediction. Background Technology

[0002] Wind energy, as an important clean energy source, is being developed and utilized more and more widely. However, due to the influence of complex environmental uncertainties, wind power output inherently possesses randomness, volatility, and intermittency, resulting in a high degree of uncertainty in wind power generation. This uncertainty poses a severe challenge to the stable grid connection of wind farms and the power grid system; if the changes in wind power output cannot be accurately predicted and the grid operation status cannot be adjusted in a timely manner, the system's safety margin will decrease, seriously threatening the stable operation of the grid. Therefore, high-precision wind power forecasting is the core foundation for ensuring the stable operation of the power system and promoting wind power consumption.

[0003] Ultra-short-term (USST) wind power forecasting aims to predict wind power output within the next few tens of minutes, serving as a crucial data foundation for real-time grid control. Traditional deep learning-based forecasting methods typically use historical power generation of individual wind turbines and meteorological data as input, predicting individual turbine power using a neural network model and then aggregating the data to obtain the forecast for the entire wind farm. However, historical power data is significantly affected by complex meteorological conditions and the coupling interference of wind field states among multiple turbines within the farm, often containing abrupt and irregular non-stationary variations. This makes it difficult for models to effectively extract the true mapping relationship between actual power generation and wind speed meteorological data when faced with variable data patterns, thus reducing the accuracy of USST wind power forecasting under complex operating conditions. Summary of the Invention

[0004] In a first aspect, embodiments of this application provide an ultra-short-term wind power prediction method, which includes the following steps: Real-time acquisition of wind speed data and active power of each wind turbine in the wind power plant; Based on the spatial distribution of wind turbines and historical wind speed data, wind turbines are clustered to obtain clusters with similar wind conditions. According to the difference in active power of each wind turbine at any given time and its adjacent time, the local jump index of each wind turbine at any given time is constructed. In any cluster with similar wind conditions, the deviation of the local jump index of each wind turbine from the local jump index of all other wind turbines within a preset historical operating cycle is analyzed to construct the abnormal fluctuation evaluation index of each wind turbine. By analyzing the fluctuation differences between wind speed data and active power of each wind turbine during the historical operating cycle, and combining the abnormal fluctuation evaluation index, the evaluation coefficient of each wind turbine during the historical operating cycle is determined. The training loss weights for each wind turbine are determined based on the evaluation coefficients and used to train the ultra-short-term wind power prediction model to predict the wind power of each wind turbine.

[0005] Preferably, the step of clustering wind turbines to obtain clusters based on similar wind conditions includes: Use all wind turbines as input to the clustering algorithm, and output all clusters; Within each cluster, based on the difference in wind speed data between any two wind turbines, all wind turbines in each cluster are further divided. By traversing all clusters, the final wind turbine cluster is denoted as the same wind condition cluster.

[0006] Preferably, the parameters in the clustering process are set as follows: the metric distance is set to the Euclidean distance between the three-dimensional spatial coordinates of the wind turbine, the search radius is a preset multiple of the wind turbine rotor diameter, and the elevation threshold is the wind turbine rotor diameter.

[0007] Preferably, the local jump index of each wind turbine at any given time is the average of the differences in active power between each wind turbine at any given time and all its adjacent times.

[0008] Preferably, the process for constructing the abnormal fluctuation evaluation index for each wind turbine is as follows: In any cluster with the same wind conditions, the difference between the maximum local jump index of each wind turbine in the historical operating cycle and the mean of the maximum local jump index of all other wind turbines is used as the abnormal fluctuation evaluation index of each wind turbine in the historical operating cycle of any cluster with the same wind conditions.

[0009] Preferably, the evaluation coefficient of each wind turbine in the historical operating cycle is positively correlated with the fluctuation difference between the actual active power and the theoretical active power of each wind turbine in the historical operating cycle and the abnormal fluctuation evaluation index.

[0010] Preferably, the training loss weight of each wind turbine is negatively correlated with the evaluation coefficient.

[0011] Preferably, the training of the ultra-short-term wind power prediction model includes: Historical wind direction data for each wind turbine is obtained and used as input to the neural network model along with historical wind speed data. Historical active power is used as output to the neural network model. The training loss weights of each wind turbine are used as weight parameters of the corresponding loss values ​​in the neural network model for model training. The trained model is denoted as the ultra-short-term wind power prediction model.

[0012] Secondly, embodiments of this application provide an ultra-short-term wind power prediction device, wherein the device stores a computer program, and when the computer program is executed by a processor, it implements an ultra-short-term wind power prediction method as described in any of the above claims.

[0013] Thirdly, embodiments of this application also provide an ultra-short-term wind power prediction system, including a memory, a processor, and a computer program stored in the memory and running on the processor. When the processor executes the computer program, it implements the steps of any of the above-described ultra-short-term wind power prediction methods.

[0014] As can be seen from the above embodiments, the ultra-short-term wind power prediction method provided in this application has at least the following beneficial effects: This application selects clusters of wind conditions driven by similar wakes and micro-meteorological conditions as dynamic reference baselines, compares the local jump in single-unit power with the group fluctuation level, effectively separates normal wind condition fluctuations from abnormal interference, quantifies the significance and credibility of abnormal fluctuations in historical data, and lays the foundation for subsequent removal of abnormal feature interference and guiding the model to focus on learning high-quality data. Furthermore, this application uses cross-correlation analysis to remove the response lag interference caused by wind turbine inertia, quantifies the degree of decoupling between wind speed and active power, and integrates it with indicators that characterize time-series anomalous jumps. It comprehensively evaluates data quality from two dimensions: internal fluctuation anomaly and external causal mismatch, providing a precise basis for weight allocation to suppress interference from inferior data and guide the model to focus on high-quality feature learning in subsequent model training. Ultimately, this application transforms the evaluation coefficients of data quality into training loss weights. During model backpropagation, high-quality normal data is assigned high weights to deeply fit its physical laws, while low-quality abnormal data is assigned low weights to suppress its interference. This effectively avoids overfitting of the model to abnormal features, ensures that the model accurately learns the true mapping relationship between wind conditions and power, and improves the accuracy of ultra-short-term wind power prediction. Attached Figure Description

[0015] Figure 1 A flowchart illustrating the steps of an ultra-short-term wind power prediction method provided in one embodiment of this application; Figure 2 This is a schematic diagram of the evaluation coefficient extraction process provided in one embodiment of this application. Detailed Implementation

[0016] The following description, in conjunction with the accompanying drawings, details the specific scheme of the ultra-short-term wind power prediction method, device, and system provided in this application.

[0017] Please see Figure 1 The diagram illustrates a flowchart of an ultra-short-term wind power prediction method according to an embodiment of this application. The method includes the following steps: S1: Real-time acquisition of wind speed data and active power of each wind turbine in the wind power plant.

[0018] The system acquires wind speed, wind direction, active power, and three-dimensional spatial coordinates of each wind turbine in the wind farm. Wind speed data is collected in real time by wind sensors on the wind turbine nacelle; wind direction data is predicted by the local meteorological station; and active power is collected by the distribution cabinet at the wind turbine end.

[0019] It should be noted that in this embodiment, the data acquisition frequency is set to f, which is 0.1Hz. In actual applications, as other implementation methods, implementers can also set it according to specific circumstances. This embodiment does not impose any special restrictions.

[0020] S2: Based on the spatial distribution of wind turbines and historical wind speed data, wind turbines are clustered to obtain clusters with similar wind conditions; based on the difference in active power of each wind turbine at any given time and its adjacent time, a local jump index for each wind turbine at any given time is constructed; in any cluster with similar wind conditions, the deviation of the local jump index of each wind turbine from the local jump index of all other wind turbines within a preset historical operating cycle is analyzed to construct an abnormal fluctuation evaluation index for each wind turbine.

[0021] The wake effect is prevalent in wind farms. The obstruction and dissipation of airflow by the front-row wind turbines significantly reduces the actual incoming wind speed for the rear-row turbines, resulting in substantial differences in the actual operating wind speed at different spatial locations. This greatly increases the uncertainty of wind turbine power generation variations within the farm. Therefore, to accurately assess the reliability of wind turbine operating data, it is not possible to analyze all wind turbines in isolation. Instead, it is necessary to find wind turbines driven by similar wind conditions as a reference benchmark. This embodiment uses the spatial distribution of wind turbines and historical wind speed data to cluster them into groups based on similar wind conditions. The specific process is as follows: In this embodiment, all wind turbines are used as input to the clustering algorithm. The metric distance is set as the Euclidean distance between the three-dimensional spatial coordinates of the wind turbines, the search radius is a preset multiple of the wind turbine rotor diameter, the elevation threshold is the wind turbine rotor diameter, and the final output is all clusters. It should be noted that the preset value of 3 in this embodiment is set based on the empirical attenuation law and spatial correlation of the wake effect between wind turbines. In a wind farm, the wake effect generated by the front wind turbines usually reduces the wind speed most significantly within a range of 3 to 5 times the rotor diameter downstream. Beyond this distance, the wake effect gradually dissipates and tends to the background wind speed. Therefore, setting the value to 3 times the rotor diameter can more likely classify wind turbines that are directly affected by the same wake and have a strong spatial wind condition correlation into the same initial cluster. This avoids the problem of setting the value too small, which would cause turbines belonging to the same wake influence domain to be abruptly separated, and also prevents the problem of setting the value too large, which would incorrectly merge distant turbines with drastically different wind conditions. Implementers can make adaptive adjustments based on the atmospheric stability, terrain roughness, and turbine spacing of the actual wind farm.

[0022] Furthermore, within each cluster, based on the difference between the wind speed data of any two wind turbines, all wind turbines in each cluster are further divided. Specifically: within the historical operating time of the wind turbines, a preset number of time periods are randomly selected, and the DTW distance between the wind speed of each wind turbine and the wind speed of the other wind turbines in any given time period is calculated. The average of the DTW distances within the preset number of time periods is taken as the difference degree between each wind turbine and the other wind turbines. Further, the mean and standard deviation of the difference degree between all wind turbines in each cluster are calculated, and the sum of the mean difference degree and one standard deviation of the difference degree is taken as the difference degree threshold. Wind turbines with a difference degree less than or equal to the difference degree threshold are grouped into a cluster. This process is repeated for all clusters, and the resulting wind turbine cluster is denoted as the "same wind condition cluster." It should be noted that since a smaller DTW distance indicates a more similar wind speed sequence, this embodiment sets the difference degree threshold to the mean plus one standard deviation. The purpose is to accommodate reasonable minor fluctuations among units operating under the same wind conditions, while effectively eliminating outlier units that are significantly different and clearly do not belong to the same wind condition pattern, ensuring that units within the same wind condition cluster have both high group similarity and a reasonable sample size.

[0023] It should be noted that in this embodiment, the preset quantity is 10, and the length of each time period is 30 minutes. Ten such independent short periods (a total of 5 hours of valid data) are selected to calculate the average DTW distance. This effectively smooths out random interference from individual occasional abnormal periods, ensuring that the calculated sequence difference truly reflects the long-term similarity of wind conditions between the two wind turbines. It also avoids computational redundancy and the introduction of excessively long-term wind condition changes due to excessively large values. The time period length is set at 30 minutes because, in the ultra-short-term wind power prediction scenario, 30 minutes corresponds precisely to a typical short period of wind condition evolution, fully encompassing the rise, fall, and transition of wind speed. Implementers can adjust the above values ​​according to the regional climate characteristics of the wind power generation site; this embodiment does not impose any special restrictions.

[0024] According to the concept of wind-condition clustering, all wind turbines within the cluster are driven by the same or highly similar micro-meteorological conditions and wake effects in terms of spatial distribution. This results in inherent group consistency and synchronicity in wind speed changes and power generation fluctuations at the same time scale. This means that when wind conditions change normally, the wind turbines within the cluster should exhibit a coordinated response pattern. If any single unit deviates from this group pattern, it can be determined that it is not caused by normal wind conditions. Wind-condition clustering essentially provides a dynamic group reference baseline with physical causality for subsequent assessment of whether the data fluctuations of a single unit are abnormal. This transforms the extraction of abnormal features from traditional isolated single-unit analysis to robust assessment based on relative group deviation.

[0025] Furthermore, in the actual operation of wind farms, the output power of wind turbines often exhibits significant randomness and fluctuations due to the combined influence of complex meteorological conditions and environmental factors. Because wind speed itself has intermittent and turbulent characteristics, coupled with wake effects and terrain disturbances, the active power of wind turbines changes over time with a non-stationary but continuous dynamic characteristic. That is, there may be large fluctuations within local short-term windows, but the overall trend is relatively stable, without isolated drastic abrupt changes. In addition, since wind turbines in a cluster share similar wind speed conditions, although the operating states of these turbines are difficult to maintain completely uniform at the same time scale due to environmental and wind field uncertainties, the degree of fluctuation should exhibit group consistency; that is, the data fluctuation degree of a single wind turbine at the same moment will not significantly deviate from the data fluctuation degree of other wind turbines in the same cluster.

[0026] Based on the above physical characteristic analysis, it can be seen that if a single wind turbine experiences a drastic jump deviating from the overall fluctuation level of the wind turbine cluster, it means that the data is not driven by normal wind conditions but has been subjected to abnormal interference. Therefore, it is necessary to quantify the degree of local jump of a single wind turbine and measure its deviation level relative to the wind turbine cluster. Thus, this embodiment constructs a local jump index for each wind turbine at any given time based on the difference in active power between each wind turbine at any given time and its adjacent times. Within any wind-condition cluster, the deviation of the local jump index of each wind turbine within a preset historical operating cycle from the local jump indices of all other wind turbines is analyzed to construct an abnormal fluctuation evaluation index for each wind turbine. The specific calculation process is as follows: In this embodiment, firstly, based on the difference in active power of each wind turbine at any given time and its adjacent times, a local jump index for each wind turbine at any given time is constructed. Specifically: In this embodiment, the average difference in active power between each wind turbine at any given moment and all its adjacent moments is denoted as the local jump index of each wind turbine at that moment. This index characterizes the severity of transient changes in the active power of the wind turbine in the local time series, reflecting the level of non-stationary fluctuations caused by abnormal interference or faults in the wind turbine's operating state. The larger the local jump index, the more drastic the power jump at that moment compared to the moments before and after, indicating that the wind turbine is very likely to have suffered abnormal shocks such as gust wind shear, sudden changes in pitch angle, or transient faults in the converter, causing the power output to deviate significantly from the normal physical trajectory. If such abnormal fluctuation data is input into the model, it will disrupt the continuity of the time series characteristics, causing the model to learn false patterns and thus reduce the prediction accuracy. Conversely, the smaller the local jump index, the smoother the power change at that moment, reflecting that the wind turbine is in a stable operating state driven by normal wind conditions. The data is real, reliable, and conforms to continuous dynamic laws, providing high-quality stable samples for the prediction model and helping the model to accurately extract the true mapping relationship between wind speed and power.

[0027] It should be noted that for the start and end times of the historical operating cycle, only the difference between adjacent times existing on one side is used for calculation.

[0028] Furthermore, this embodiment constructs an abnormal fluctuation evaluation index for each wind turbine by analyzing the deviation of its local jump index from the local jump index of all other wind turbines within a preset historical operating cycle in any wind condition cluster. Specifically: In any cluster with the same wind conditions, the difference between the maximum local jump index of each wind turbine within a preset historical operating cycle and the average of the maximum local jump index of all other wind turbines is used as the abnormal fluctuation evaluation index of each wind turbine within the historical operating cycle of any cluster with the same wind conditions.

[0029] It should be noted that the preset historical operating cycle length in this embodiment is 30 minutes because, in the ultra-short-term wind power prediction scenario, 30 minutes corresponds precisely to a typical micro-perpetual cycle of wind condition evolution within the atmospheric boundary layer. This not only fully covers the short-term fluctuations in wind speed from start to finish, ensuring that the maximum local jump index calculated within this window can truly capture the full picture of extreme wind condition fluctuations, but also avoids the awkward mixing and smoothing out of real transient changes caused by an excessively long window. It also highly matches the ultra-short-term command cycle for real-time power balance control by the power grid dispatch center. Of course, this preset length can be other values, and implementers can adaptively adjust it according to the time resolution of the specific prediction task and the local microclimate characteristics of the wind farm. This embodiment does not impose any special restrictions.

[0030] It should be noted that there are many methods for measuring the differences between data. In this embodiment, the absolute difference between the active power of each wind turbine at any given time and its adjacent times is taken as the difference between the active power of each wind turbine at any given time and its adjacent times. The absolute difference between the maximum local jump index of each wind turbine in the historical operating cycle and the mean of the maximum local jump index of all other wind turbines is taken as the difference between the maximum local jump index of each wind turbine in the historical operating cycle and the mean of the maximum local jump index of all other wind turbines. In practical applications, as other implementation methods, implementers may also choose other methods for measuring the differences between data according to specific circumstances. This embodiment does not impose any special restrictions.

[0031] The abnormal fluctuation evaluation index, as understood from the data, characterizes the significance of abnormal fluctuations in historical operating data, reflecting the reliability of data quality and the strength of abnormal interference. Since the calculation formula is the absolute difference between the maximum local jump index and the group mean, and the local jump index is derived from the active power difference, its dimensions are consistent with those of active power (e.g., MW). This is because it retains the absolute physical magnitude of power fluctuations without dimensionless processing. Its calculation is influenced by the maximum local jump index of wind turbines within the historical operating cycle and the maximum local jump index in the same wind condition cluster. The difference in the mean of the jump index has an impact. When the difference is larger, the abnormal fluctuation evaluation index is larger, indicating that the extreme fluctuation of the corresponding wind turbine far exceeds the normal level of the group, reflecting that it has suffered a serious abnormal interference unique to this turbine. The data has extremely low credibility, and if it is input into the model, it will seriously poison the model feature space and lead to prediction distortion. Conversely, when the difference is smaller, the abnormal fluctuation evaluation index is smaller, indicating that the fluctuation of the corresponding wind turbine is highly consistent with the group under the same wind conditions, reflecting that it is in a standard operating state driven by normal wind conditions. The data is real and reliable and can provide high-quality learning samples for subsequent prediction model training.

[0032] Thus, this embodiment effectively separates normal wind condition fluctuations from abnormal interference by selecting clusters of similar wind conditions driven by similar wakes and micro-meteorological conditions as dynamic reference baselines, and compares the degree of local jumps in single-unit power with the level of group fluctuations. It quantifies the significance and reliability of abnormal fluctuations in historical data, laying the foundation for subsequent removal of abnormal feature interference and guiding the model to focus on learning high-quality data.

[0033] S3: By analyzing the fluctuation differences between wind speed data and active power of each wind turbine during the historical operating cycle, and combining the abnormal fluctuation evaluation index, the evaluation coefficient of each wind turbine during the historical operating cycle is determined.

[0034] In the power generation process of wind turbines, wind speed is the core factor directly driving the output power of the wind turbine, and the two have a clear coupling relationship in dynamic changes. However, as a power generation facility with large inertia, the output power of a wind turbine has an inherent lag in response to changes in wind speed; despite this, the two still have a strong correlation in terms of their changing trends and magnitudes. When wind speed changes significantly at the same time scale, but the power generation does not show the same fluctuation, it indicates that the coupling consistency between the two has been disrupted. This is often due to abnormal interference or abnormal unit operating conditions, resulting in abnormal operating characteristics in the data.

[0035] Based on the above analysis, to accurately identify the degree of disruption of this coupling relationship, the response lag effect caused by large inertia must first be eliminated; otherwise, misaligned timing deviations will be misjudged as inconsistent fluctuations, thus masking the true abnormal characteristics. Therefore, this embodiment needs to analyze the cross-correlation between the active power and theoretical active power of each wind turbine within the historical operating cycle to obtain the response lag delay, align the wind speed data and active power in time according to the response lag delay, analyze the fluctuation difference between the aligned wind speed data and active power, and combine it with the abnormal fluctuation evaluation index to determine the evaluation coefficient of each wind turbine within the historical operating cycle. The specific process is as follows: In this embodiment, wind speed-power curves for each wind turbine are constructed based on wind turbine size parameter data. The specific process is as follows: the swept area of ​​the wind turbine is calculated based on the rotor diameter and air density in the wind turbine size parameter data. The wind energy capture efficiency is determined by combining the Bates limit theory. The theoretical active power that the wind turbine can output at different wind speeds is calculated using the wind energy conversion formula. Then, the theoretical active power is segmented and corrected by combining the cut-in wind speed, rated wind speed, and cut-out wind speed parameters in the wind turbine size parameters. Finally, the wind speed-power curve and its corresponding function, which characterize the mapping relationship between the input wind speed and the output theoretical active power of the wind turbine model, are generated by fitting discrete data points. Thus, by substituting the wind speed at each moment into the corresponding function of the wind speed-power curve, the theoretical active power at each moment can be obtained.

[0036] Furthermore, based on the cross-correlation between the active power and theoretical active power of each wind turbine during the historical operating cycle, the response lag delay is obtained. Specifically, the normalized values ​​of the active power and theoretical active power of each wind turbine during the historical operating cycle are substituted into the cross-correlation function. In the non-zero time delay, the time delay corresponding to the maximum cross-correlation coefficient is taken as the response lag delay. The calculation process of the cross-correlation coefficient is a well-known technique and will not be elaborated further.

[0037] It should be noted that, when calculating the cross-correlation coefficient, since wind speed and active power have different dimensions, in order to eliminate the influence of data dimensions on the calculation of the cross-correlation coefficient, this embodiment uses the maximum-minimum normalization method to normalize the wind speed data and active power respectively. Specifically, the maximum and minimum values ​​of all wind speed data within the historical operating period are used as the maximum and minimum value parameters in the maximum-minimum normalization function. Similarly, the maximum and minimum values ​​of all active power within the historical operating period are used as the maximum and minimum value parameters in the maximum-minimum normalization function. In practical applications, as other implementation methods, implementers may also choose other normalization methods such as z-score normalization according to specific circumstances. This embodiment does not impose any special restrictions on the selection of normalization methods.

[0038] The process of normalizing data using the maximum-minimum normalization method is a well-known technique and will not be elaborated further.

[0039] Furthermore, within the historical operating cycle, starting from the initial moment, the active power within the preceding response lag time delay is removed, and ending at the final moment, the wind speed data within the following response lag time delay is removed, thus aligning the wind speed data and active power in time sequence. To eliminate the interference of the inherent nonlinear cubic relationship between wind speed and active power on distance calculation, the wind speed data at each time point after time sequence alignment is input into the wind speed-power curve constructed in step S3 to obtain the theoretical active power corresponding to each time point. The Euclidean distance between the normalized value of the aligned actual active power and the normalized value of the theoretical active power is calculated. For ease of expression, this distance is denoted as the fluctuation difference index of each wind turbine within the historical operating cycle.

[0040] The fluctuation difference index characterizes the degree to which the actual power generation of a wind turbine deviates from its theoretical physical response trajectory. It reflects whether the actual electromechanical conversion process of the wind turbine follows normal physical laws and is a dimensionless index. Its calculation is affected by the difference between the actual active power and the theoretical active power. Since the wind speed is mapped to theoretical power through the wind speed-power curve, the inherent mathematical differences of the nonlinear cubic relationship are eliminated. The larger the difference, the larger the fluctuation difference index, indicating that the actual power deviates significantly from the theoretical expectation under the same wind speed. This reflects serious decoupling phenomena such as pitch angle faults, generator disconnection from the grid, or converter abnormalities in the wind turbine. Inputting such data into the model will prevent the model from learning the correct causal mapping, rendering the prediction results meaningless. Conversely, a smaller fluctuation difference index indicates a high degree of overlap between the actual power and the theoretical expectation trajectory, reflecting a sensitive and healthy wind turbine response. The corresponding data allows the model to accurately capture the true physical coupling law between wind speed and power, thereby improving prediction reliability.

[0041] Furthermore, this embodiment constructs evaluation coefficients for each wind turbine within its historical operating cycle based on the fluctuation difference index and the abnormal fluctuation evaluation index, specifically as follows: In this embodiment, the evaluation coefficient of each wind turbine in the historical operating cycle is positively correlated with the fluctuation difference between the wind speed data and active power of each wind turbine in the historical operating cycle (i.e., the fluctuation difference index) and the abnormal fluctuation evaluation index.

[0042] Preferably, the schematic diagram of the evaluation coefficient extraction process provided in this embodiment is as follows: Figure 2 As shown.

[0043] It should be understood that a positive correlation means that the dependent variable increases as the independent variable increases, and the dependent variable decreases as the independent variable decreases. The specific relationship can be additive or multiplicative, etc., and is determined by the actual application. This application does not impose any special restrictions.

[0044] Preferably, as an implementation method, in this embodiment, the normalized value of the fluctuation difference index of each wind turbine and the normalized value of the abnormal fluctuation evaluation index within the historical operating cycle are recorded as the evaluation coefficient of each wind turbine within the historical operating cycle.

[0045] It should be noted that, since the fluctuation difference index and the abnormal fluctuation evaluation index differ in magnitude, this embodiment uses the maximum-minimum value normalization method to normalize both the fluctuation difference index and the abnormal fluctuation evaluation index in order to eliminate this difference. When normalizing the fluctuation difference index, the maximum and minimum values ​​of the fluctuation difference indices of all wind turbines within the historical operating cycle are used as the maximum and minimum value parameters in the maximum-minimum value normalization method, respectively. Similarly, the maximum and minimum values ​​of the abnormal fluctuation evaluation indices of all wind turbines are used as the maximum and minimum value parameters in the maximum-minimum value normalization method, respectively. In practical applications, as other implementation methods, implementers may also use other normalization methods according to specific circumstances. This embodiment does not impose any special restrictions on the selection of normalization methods.

[0046] Based on the evaluation coefficient, it can be understood that the evaluation coefficient is a dimensionless index, and its calculation is affected by two factors: the abnormal fluctuation evaluation index and the fluctuation difference index. The former reflects the abnormality of the time series jumps within the data of the wind turbine, and the latter reflects the abnormality of the causal response of the data input and output. When these two factors are larger, the evaluation coefficient is larger, indicating that the data of the corresponding wind turbine not only has abnormal jumps in time series, but also has serious decoupling in physical response. This reflects that the data within the wind turbine is a low-quality sample contaminated by multiple interferences. If it is not suppressed, the model will learn incorrect pseudo-laws, leading to a sharp increase in prediction error. Conversely, when these two factors are smaller, the evaluation coefficient is smaller, indicating that the data within the corresponding wind turbine is both stable and conforms to the physical coupling law, reflecting that it is a high-quality normal sample that can provide the purest feature paradigm for the model and help improve the prediction accuracy of the model.

[0047] Thus, this embodiment has eliminated the response lag interference caused by wind turbine inertia through cross-correlation analysis, quantified the degree of decoupling between wind speed and active power fluctuations, and integrated it with indicators characterizing time-series anomalous jumps. It comprehensively evaluated data quality from two dimensions: internal fluctuation anomalies and external causal mismatch, providing a precise basis for weight allocation to suppress interference from inferior data and guide the model to focus on high-quality feature learning in subsequent model training.

[0048] S4: Determine the training loss weight of each wind turbine based on the evaluation coefficient, and use it to train the ultra-short-term wind power prediction model to predict the wind power of each wind turbine.

[0049] In traditional neural network model training, all input training samples are typically assigned the same loss weight, ensuring consistent learning levels for both normal and abnormal data. However, as analyzed in the preceding steps, historical operational data from wind farms inevitably contains abnormal fluctuations caused by anomalous disturbances or abnormal unit conditions. If such data occupies the regular weights during training, its anomalous characteristics will be forcibly fitted by the model, leading to decreased generalization ability and reduced prediction accuracy for normal wind conditions. Therefore, it is necessary to incorporate the data quality evaluation results extracted in the aforementioned steps into the model training process, enabling the model to focus on learning from high-quality data.

[0050] Based on the above analysis, this embodiment determines the training loss weight of each wind turbine based on the evaluation coefficient, which is used to train the ultra-short-term wind power prediction model to predict the wind power of each wind turbine. That is, by converting the evaluation coefficient reflecting the degree of data anomaly into a loss penalty factor during model training, the weight of data with a high degree of anomaly in the total loss is reduced, thereby suppressing the interference of abnormal features. The specific process is as follows: In this embodiment, firstly, the training loss weights of each wind turbine within the historical operating cycle are determined based on the evaluation coefficients. Specifically: In this embodiment, the training loss weight of each wind turbine is negatively correlated with the evaluation coefficient.

[0051] It should be understood that a negative correlation means that the dependent variable decreases as the independent variable increases, and the dependent variable increases as the independent variable decreases. It can be a subtraction relationship or a division relationship, depending on the actual application.

[0052] Preferably, as one implementation method, the training loss weights of wind turbine i during the historical operating cycle in this embodiment are... The expression is: In the formula, This represents the evaluation coefficient of wind turbine i within the historical operating cycle.

[0053] Based on the training loss weight, it can be understood that when the neural network model performs backpropagation parameter updates, the proportion coefficient assigned to the prediction error of a single training sample in the total loss function is essentially a penalty mechanism in the model optimization process, which transforms the evaluation result of data quality into a penalty mechanism. The training loss weight is negatively correlated with the evaluation coefficient, which means that high-quality (low evaluation coefficient) normal data is given a large weight, so that the model pays close attention to and deeply fits the true physical laws of these samples during training. On the other hand, low-quality (high evaluation coefficient) abnormal data is given a very small weight, so that the model almost ignores the interference of its abnormal features when updating parameters. This guides the model to avoid the data traps of abnormal fluctuations and coupling mismatch on a macroscopic level, ensuring that the model always converges along the direction of normal wind power operation characteristics, and finally achieves high-precision prediction of ultra-short-term wind power.

[0054] Furthermore, each wind turbine within any historical operating cycle is taken as a sample. The input feature vector corresponding to this sample is composed of the wind speed and wind direction data obtained in step S1, and its dimension is... ( (This represents the number of sampling points within the historical operating cycle). The label vector of this sample is the corresponding active power, with dimensions of... This allows us to obtain the wind power clusters within the entire wind farm under the same wind conditions. A total sample is defined, where I represents the number of all wind turbines in the same wind condition cluster, and N represents the number of all historical operating cycles. In this embodiment, the total sample is set to 100. Further, the total sample is randomly divided into a training set and a test set according to a preset ratio (e.g., 8:2). Further, during the training phase, the training set samples are used as input to a Long Short-Term Memory (LSTM) neural network model. The Adam optimizer and a weighted MSE (mean squared error) loss function are used for iterative optimization. The loss value of each sample is multiplied by its corresponding training loss weight, and the multiplication results for all samples are accumulated into the total loss. This reduces the tendency of the neural network model to fit abnormal fluctuations in the data during backpropagation parameter updates. When the loss function value on the validation set no longer decreases for a preset number of consecutive rounds, an early stop mechanism is triggered to end the training. The accuracy of the trained model is verified using the test set. Once the accuracy is achieved, it is determined to be an ultra-short-term wind power prediction model.

[0055] After obtaining the ultra-short-term wind power prediction model, the numerical meteorological forecast data (including forecast wind speed and forecast wind direction) of the target wind turbine in the future ultra-short-term time period is used as the model input. The feature dimensions and arrangement order of the input data are strictly consistent with the input samples in the model training stage. The forecast data is input into the trained ultra-short-term wind power prediction model. The model uses its internal network structure to perform forward propagation calculation on the multivariate time series features and directly outputs the active power of the target wind turbine at each sampling time in the future ultra-short-term time period.

[0056] The process of training the Long Short-Term Memory (LSTM) neural network model is a well-known technique and will not be described in detail here.

[0057] Thus, this embodiment transforms the evaluation coefficient of data quality into training loss weights. During model backpropagation, high-quality normal data is given high weights to deeply fit its physical laws, while low-quality abnormal data is given low weights to suppress its interference. This effectively avoids the model overfitting abnormal features, ensures that the model accurately learns the true mapping relationship between wind conditions and power, and improves the accuracy of ultra-short-term wind power prediction.

[0058] Based on the same inventive concept as the above method, this application also provides an ultra-short-term wind power prediction device, wherein the device stores a computer program, and when the computer program is executed by a processor, it implements an ultra-short-term wind power prediction method as described above.

[0059] Based on the same inventive concept as the above method, this application embodiment also provides an ultra-short-term wind power prediction system, including a memory, a processor, and a computer program stored in the memory and running on the processor. When the processor executes the computer program, it implements the steps of any one of the above-described ultra-short-term wind power prediction methods.

Claims

1. A method for predicting ultra-short-term wind power, characterized in that, The method includes the following steps: Real-time acquisition of wind speed data and active power of each wind turbine in the wind power plant; Based on the spatial distribution of wind turbines and historical wind speed data, wind turbines are clustered to obtain clusters with similar wind conditions. According to the difference in active power of each wind turbine at any given time and its adjacent time, the local jump index of each wind turbine at any given time is constructed. In any cluster with similar wind conditions, the deviation of the local jump index of each wind turbine from the local jump index of all other wind turbines within a preset historical operating cycle is analyzed to construct the abnormal fluctuation evaluation index of each wind turbine. By analyzing the fluctuation differences between wind speed data and active power of each wind turbine during the historical operating cycle, and combining the abnormal fluctuation evaluation index, the evaluation coefficient of each wind turbine during the historical operating cycle is determined. The training loss weights for each wind turbine are determined based on the evaluation coefficients and used to train the ultra-short-term wind power prediction model to predict the wind power of each wind turbine.

2. The ultra-short-term wind power prediction method as described in claim 1, characterized in that, The clustering of wind turbines to obtain clusters with similar wind conditions includes: Use all wind turbines as input to the clustering algorithm, and output all clusters; Within each cluster, based on the difference in wind speed data between any two wind turbines, all wind turbines in each cluster are further divided. By traversing all clusters, the final wind turbine cluster is denoted as the same wind condition cluster.

3. The ultra-short-term wind power prediction method as described in claim 2, characterized in that, The parameters for the clustering process are set as follows: the metric distance is set to the Euclidean distance between the three-dimensional spatial coordinates of the wind turbine, the search radius is a preset multiple of the wind turbine rotor diameter, and the elevation threshold is the wind turbine rotor diameter.

4. The ultra-short-term wind power prediction method as described in claim 1, characterized in that, The local jump index of each wind turbine at any given time is the average of the differences in active power between each wind turbine at any given time and all its adjacent times.

5. The ultra-short-term wind power prediction method as described in claim 1, characterized in that, The process for constructing the abnormal fluctuation evaluation index for each wind turbine is as follows: In any cluster with the same wind conditions, the difference between the maximum local jump index of each wind turbine in the historical operating cycle and the mean of the maximum local jump index of all other wind turbines is used as the abnormal fluctuation evaluation index of each wind turbine in the historical operating cycle of any cluster with the same wind conditions.

6. The ultra-short-term wind power prediction method as described in claim 1, characterized in that, The evaluation coefficients of each wind turbine during the historical operating cycle are positively correlated with the fluctuation difference between the actual active power and the theoretical active power of each wind turbine during the historical operating cycle, as well as the abnormal fluctuation evaluation index.

7. The ultra-short-term wind power prediction method as described in claim 1, characterized in that, The training loss weights and evaluation coefficients of each wind turbine are negatively correlated.

8. The ultra-short-term wind power prediction method as described in claim 1, characterized in that, The training of the ultra-short-term wind power prediction model includes: Historical wind direction data for each wind turbine is obtained and used as input to the neural network model along with historical wind speed data. Historical active power is used as output to the neural network model. The training loss weights of each wind turbine are used as weight parameters of the corresponding loss values ​​in the neural network model for model training. The trained model is denoted as the ultra-short-term wind power prediction model.

9. A short-term wind power prediction device, wherein the device stores a computer program, characterized in that, When the computer program is executed by the processor, it implements the ultra-short-term wind power prediction method as described in any one of claims 1-8.

10. An ultra-short-term wind power forecasting system, comprising a memory, a processor, and a computer program stored in the memory and running on the processor, characterized in that, When the processor executes the computer program, it implements the ultra-short-term wind power prediction method as described in any one of claims 1-8.

Images