Wind power cluster output prediction method and system based on spatial-temporal characteristic coupling

Abstract

The invention discloses a spatio-temporal feature coupled wind power cluster output prediction method and a spatio-temporal feature coupled wind power cluster output prediction system. Firstly, according to valid data of missing values in time and space neighborhoods, the missing values are classified and processed, so that the integrity and availability of the data are improved; secondly, identifying upstream and downstream relationships between each fan and other fans at each moment by mining space-time information among the fans in a wind power cluster, and constructing leading features through a wind propagation effect, and taking the leading features as alternative features for predicting output conditions of corresponding downstream fans; thirdly, evaluating the influence of the alternative features on wind power prediction in multiple aspects, and screening out features which contribute significantly to prediction of wind power cluster output; and finally, constructing a multi-input multi-output model structure, inputting the screened effective features into a deep learning model at one time, and outputting a wind power cluster output prediction result. The model provided by the invention is obviously superior to a reference model in evaluation indexes, and shows excellent prediction performance and reliability.

Description

Technical Field

[0001] The present invention relates to a method and system for predicting the output of a wind power cluster by coupling spatio-temporal characteristics, which can significantly improve the accuracy of wind power prediction and belongs to the technical field of wind power prediction. Background Art

[0002] With the increasing global demand for clean energy, wind energy, as a clean and sustainable energy form, is of great significance for reducing pollution and promoting environmental sustainability. However, the random and intermittent characteristics of wind energy pose significant challenges to the effective scheduling and utilization of wind power. To address these issues, accurate wind power prediction is crucial, which helps to optimize wind power scheduling, improve wind energy utilization rate, and alleviate the phenomenon of wind curtailment, thereby achieving the economic and efficient operation of the power system.

[0003] Currently, the main challenges faced by wind power prediction lie in its intermittency and uncertainty, which make it extremely difficult to obtain accurate leading information. Without leading information, relying solely on the past output of individual wind turbines to predict future output, it is difficult to further improve the prediction accuracy. To solve this problem, existing wind power prediction models attempt to construct features with leading information and input them into the prediction model to improve the prediction effect. However, the existing feature construction methods have the following three main defects. The first is the failure to consider the small-scale interaction and influence between the output of each wind turbine within the wind power cluster. Many existing prediction models only study wind power as a time series, only focus on the data of individual wind turbines, and construct leading features using large-scale meteorological factors, while ignoring the small-scale interaction between wind turbines within the wind power cluster. For example, when the wind blows through a wind turbine, the wind speed received by other wind turbines located downstream of this wind turbine will decrease, and this phenomenon is called the wake effect or blockage effect. Such small-scale spatial information is often difficult to be captured in meteorological factors and is therefore often ignored in existing methods. The second defect is that the constructed features are relatively abstract and it is difficult to clarify their real meanings, so it is also impossible to make further improvements on the abstract features. Currently, many artificial intelligence-based prediction models usually collect a large amount of high-dimensional data and use black-box technologies such as autoencoders to reduce their dimensions when constructing input features. The low-dimensional information extracted therefrom is not only difficult to intuitively understand and unable to clarify the specific physical or real meanings it represents, but also there are great difficulties in evaluating the effectiveness and theoretical analysis of the abstract features. Third, the existing methods lack a process for systematically evaluating and screening the effectiveness of features, and often directly import large-scale meteorological factors as inputs into the prediction model. However, if the input features are not reasonably screened, the model may be interfered by redundant features or features with weak correlations, thus affecting the accuracy of the prediction results and causing large deviations in the model output. Summary of the Invention

[0004] Object of the Invention: Aiming at the problems and deficiencies existing in the prior art, the present invention provides a wind power cluster output prediction method and system with spatio-temporal feature coupling. By mining the spatio-temporal relationships among various wind turbines in the wind power cluster, the present invention constructs impact factors that can reflect the output changes of various wind turbines in the wind power cluster in advance. These impact factors can not only be clearly explained theoretically but also be easily understood in empirical analysis, providing a more intuitive and interpretable feature construction scheme. In addition, through a comprehensive feature selection method, the system evaluates the correlation between all alternative features (the constructed leading features, historical wind power data, meteorological factors, and wind turbine operating conditions) in the wind power cluster and the wind turbine output, as well as their contributions to the prediction. On this basis, irrelevant and redundant features are removed, and the key features that contribute most to the wind power cluster output prediction are selected and used as input variables for prediction models such as deep learning, thereby significantly improving the accuracy and reliability of the wind power cluster output prediction.

[0005] Technical Solution: A wind power cluster output prediction method with spatio-temporal feature coupling includes the following steps: Step 1: Collect data. The data includes wind power data, meteorological factor data, and wind turbine operating condition data, and the time of different data is aligned. The wind power data is the target sequence, and the meteorological factor data and wind turbine operating condition data are feature data.

[0006] Step 2: Spatio-temporal interpolation. The missing values in the target sequence and feature data are classified and processed.

[0007] Step 3: Construct leading indicators. Taking the deep learning model as the wind power cluster output prediction model, construct effective leading indicators as one of the input features of the prediction model. By mining the spatio-temporal information among various wind turbines in the wind power cluster, identify the upstream and downstream relationships of each wind turbine with other wind turbines at each moment, and determine the upstream wind turbines by maximizing the wind propagation effect. Use all the first-order lag historical data of the identified upstream wind turbines as leading indicators, as alternative features for predicting the output of the corresponding downstream wind turbines.

[0008] Step 4: Feature Selection. Evaluate the relationships between wind power and various alternative features, and screen out the features that are useful for predicting the output of the wind power cluster. The alternative features are mainly divided into four categories, namely historical wind power data, meteorological factors (such as wind speed, wind direction, temperature, etc.), fan operating conditions (such as nacelle direction, blade angle, etc.), and leading indicators (historical data of the upstream fans corresponding to the target fan). Three methods are used to examine the linear relationship, monotonic relationship, and complex non-linear relationship between wind power and alternative features respectively. These methods are Pearson correlation coefficient, Spearman correlation coefficient, and pre-experiment. The Pearson correlation coefficient is used to evaluate the strength and direction of the linear relationship between variables, while the Spearman correlation coefficient focuses on evaluating the monotonic association between variables, that is, the relationship where two variables are consistent in the changing trend but not necessarily linear. In order to further reveal the non-linear dependence relationships that may not be fully captured by these traditional statistical methods, the present invention adopts a pre-experiment based on random forest. This method measures the importance of features by calculating the degree of average impurity reduction of each feature in all decision trees, so as to highlight those features that can most effectively split the data for prediction. Considering the results of these three methods comprehensively, screen out the effective features that are most important for wind power cluster output prediction, thus providing a solid foundation for subsequent research.

[0009] Step 5: Use a deep learning model for multi-step prediction. Construct a multi-input multi-output model structure, input the selected effective features into the deep learning model at one time, and directly output the predicted results of the wind power cluster output at multiple future time points within a set time period (such as: from 15 minutes to 16 hours). In the initial stage, the internal parameters of the model are randomly initialized; subsequently, by minimizing the loss function (i.e., mean square error), use the backpropagation algorithm to continuously adjust the model parameters, and finally obtain a wind power output prediction model with optimized performance and high prediction accuracy.

[0010] In the said Step 1, the meteorological factors include wind speed, wind direction, and temperature; the fan operating conditions include nacelle direction and blade angle.

[0011] In the said Step 2, for the missing values in the target sequence and feature data, classify and process them, and the specific contents are as follows: Step 2.1: Preliminary screening and classification of missing values. For each missing value, first determine its neighboring observation points in the spatio-temporal domain. Based on the spatial distance, select two fans with the closest physical distance and valid observations as candidate points. At the same time, for the time dimension, define a time window and select the two time points closest to the missing moment before and after the target fan and with valid observations as candidates.

[0012] Step 2.2: Apply corresponding interpolation strategies to process the missing values according to the situation.

[0013] (a) Bilinear interpolation: When the candidate point contains both valid spatial neighboring observations and data from two timestamps before and after, the weighted bilinear interpolation method is used for filling. This method selects neighboring valid observation points in both the time and space dimensions and obtains the estimated value of the target point through weighted fusion. The calculation formula is as follows: x(·)=w s *( w s1 * x s1 + w s2 * x s2 )+ w t *( w t1 * x t1 + w t2 * x t2 ), where x(·) is the value used to fill the missing value. x s1 and x s2 respectively represent the observed values of the spatial candidate point, and x t1 and x t2 represent the observed values of the time candidate point. w s and w t are the weighted coefficients for the overall space and time dimensions respectively, and satisfy w s +w t =1. To highlight the dominant role of time continuity in the change process, a higher weight is given to the time dimension (i.e., w s <w t ), thereby improving the rationality of the interpolation result in space-time. In addition, for the observed values of the spatial candidate point, a decreasing weight is assigned according to the physical distance. Using d1 and d2 to represent the distances to the two nearest spatial candidate points respectively, the spatial weight w s1 = d1 / (d1+d2), w s2 = d2 / ( d1+d2), thus ensuring that the spatial candidate points closer in distance have a greater weight.

[0014] (b) Temporal linear interpolation: If more than a set proportion (such as more than 80%) of the data in the spatial candidate points is missing, linear interpolation is performed only relying on the data of the two valid timestamps before and after within the selected time window. This method assumes that the time series data changes smoothly in a short period and is applicable to the situation where the data is continuous in time but insufficient in spatial information.

[0015] (c) Spatial nearest neighbor interpolation: When only spatial neighboring points are available, that is, when no valid timestamp data can be obtained, the two closest observations are selected as candidate points from the spatial dimension. Here, a linear interpolation method is adopted. Different from temporal linear interpolation, it assigns weights based on spatial distance, with observations closer in distance having greater weights. This method is applicable to cases where the spatial distribution is relatively uniform and temporal information is completely missing.

[0016] (d) Data deletion: For those cases where there are neither suitable spatial neighboring points nor neighboring observations on the time axis, it is considered that the information of these points is insufficient to support reliable interpolation, so they are selected to be removed.

[0017] In addition, it is necessary to specifically explain the weight design principle. The most fundamental principle is that the sum of all weights during interpolation is 1, whether in the temporal dimension or the spatial dimension. And when implementing bilinear interpolation, the importance of the temporal dimension is particularly emphasized because it represents the results of continuous observations of the same individual. In contrast, although the spatial dimension can provide additional information, it depends on the similarity between different individuals, so its weight is set relatively low. At the same time, within the spatial dimension, the principle of "the closer the distance, the greater the correlation" is also followed, ensuring that the spatial weight gradually decreases as the distance increases.

[0018] In Step 3 described above, the wind propagation mechanism is defined as follows: Suppose when a gust of wind blows from wind turbine s1 to another wind turbine s2, within the propagation time h minutes, the wind speed of wind turbine s1 at time t is equal to the wind speed of wind turbine s2 at time t + h. This is called the wind propagation mechanism. Based on this assumption, the wind propagation effect, that is, the empirical spatio-temporal variogram g(s1, s2, h), can be formally defined as:

[0019] where x(s, t) represents the wind speed of wind turbine s at time t, and L represents the time window length. When the wind reaches wind turbine s2 from wind turbine s1 after a propagation time of h, g(s1, s2, h) = 0. Using this property to explore the upstream and downstream relationships among multiple wind turbines in a wind power cluster, identify the most effective upstream wind turbine corresponding to each wind turbine at each moment, and construct leading features using the historical data of the upstream wind turbine. The leading features constructed in this way not only capture the spatial interrelationships among multiple wind turbines in the wind power cluster, but also consider the changes in these spatial relationships at different time points, thus comprehensively reflecting the spatio-temporal characteristics of the wind power cluster output. Inputting this leading feature into the prediction model helps the model better capture the potential laws in the data through this feature, thereby improving the fitting accuracy and prediction performance of wind power output.

[0020] Step 3 described above specifically includes the following steps: Step 3.1: Determine the appropriate propagation time based on the historical power output of the wind power cluster. . Specifically, the propagation distance is obtained by multiplying the propagation time by the average wind speed. This calculated distance is roughly consistent with the actual distance between adjacent wind turbines in the wind power cluster. When the propagation time is too large, it may lead to overly coarse spatial granularity and inability to accurately capture the changes within the wind farm. In this case, the time granularity needs to be increased to improve the accuracy. On the contrary, when the propagation time is too small, the spatial granularity is too fine, which may cause the time granularity to exceed the wind farm boundary. In this case, the time granularity needs to be reduced to ensure the validity of the data.

[0021] Step 3.2: Determine the upstream wind turbines. The process of determining the upstream wind turbines of a single target wind turbine is as follows: within the time window L, calculate the empirical spatio-temporal variogram between all pairs of wind turbines . Let s2 represent the target wind turbine, and s1 traverse the remaining wind turbines. Then, select the wind turbines obtained through the formula as the upstream wind turbines of the target wind turbine at the current moment. In this way, the upstream wind turbines of a single target wind turbine can be found. By traversing all the wind turbines in the cluster as target wind turbines and repeating the above process of determining the upstream wind turbines of a single target wind turbine, the upstream wind turbines of each wind turbine can be identified. On this basis, combined with the method of rolling time window, the upstream wind turbines at different moments are further identified, so as to realize the identification of upstream wind turbines within the wind power cluster during the entire time period.

[0022] Step 3.3: Construct leading indicators. Directly use the historical data of the upstream wind turbines as leading indicators. Since the upstream and downstream relationships between the wind turbines in the wind power cluster are not static and will change with the change of wind direction, it should be noted that the upstream wind turbines at time t are identified based on the power output data of the wind power cluster from time t - L to t - L + 1.

[0023] A wind power cluster output prediction system coupling spatio-temporal features includes the following modules: Data collection module; used to collect wind power data, meteorological factor data, and wind turbine operation status data, and align the time of different data; the wind power data is the target sequence, and the meteorological factor data and wind turbine operation status data are feature data; Spatio-temporal interpolation module; classify and process the missing values in the target sequence and feature data; Leading indicator construction module; by mining the spatio-temporal information between each wind turbine in the wind power cluster, identify the upstream and downstream relationships between each wind turbine and other wind turbines at each moment, and determine the upstream wind turbines by maximizing the wind propagation effect. Use all the first-order lag historical data of the upstream wind turbines as leading indicators for alternative features to predict the output of downstream wind turbines; Feature selection module; evaluate the relationship between wind power and various alternative features, and screen out the features useful for predicting the output of a wind power cluster; The alternative features include four categories, namely historical wind power data, meteorological factors, wind turbine operating conditions, and leading indicators; three methods, namely Pearson correlation coefficient, Spearman correlation coefficient, and pre-experiment based on random forest, are used to investigate the linear relationship, monotonic relationship, and complex non-linear relationship between wind power and alternative features respectively. The union of the features with linear correlation reaching the set threshold, monotonic relationship reaching the set threshold, and random forest importance score reaching the set threshold is taken as the input feature set of the wind power cluster output prediction model; Prediction module; use a deep learning model for multi-step prediction; input the selected effective features into a multi-input multi-output deep learning model, and output the wind power cluster output prediction results at multiple future time points within a set time period.

[0024] The implementation process of the system is the same as the method, and will not be elaborated here.

[0025] A computer device, which includes a memory, a processor, and a computer program stored on the memory and executable on the processor. When the processor executes the above computer program, it implements the wind power cluster output prediction method with spatio-temporal feature coupling as described above.

[0026] Beneficial effects: Compared with the prior art, the present invention has the following advantages: (1) Accurately extract leading indicators: Use the wind propagation mechanism to distinguish upstream and downstream wind turbines and accurately obtain leading signals. By introducing the empirical spatio-temporal variation function g(s1, s2, h), upstream and downstream wind turbines can be systematically distinguished. Even in the case of fluctuations in the wind propagation process, upstream wind turbines can still be effectively identified as the leading signal source. Such accurate leading indicators not only reflect wind speed changes in advance but also solve the data drift problem and improve prediction accuracy.

[0027] (2) Effectively reduce prediction errors: Through step 4 to evaluate the relationship between wind power and alternative features, it is found that the constructed leading indicators always have a positive impact on the wind power cluster output prediction. After introducing the leading indicators into the prediction model in Step 5, whether in the ultra-short-term prediction of 15 minutes or the short-term prediction of 16 hours, the prediction errors are significantly reduced, and the overall prediction performance of the model is improved.

[0028] (3) Reduce computational costs: When constructing leading indicators in step 3, the upstream wind turbine identification process is simplified, and the algorithm complexity is reduced from O(T*S) to O(S), greatly reducing the amount of calculation while maintaining the effectiveness of upstream wind turbines. This improvement makes the application of the model in large-scale wind farms more efficient and reduces the demand for computing resources.

[0029] To verify the effectiveness of the leading indicators, extensive experiments were conducted in the present invention, including a variety of time series prediction models. The experimental results show that the leading indicators perform excellently in different scenarios, with strong robustness and adaptability, further demonstrating their value in practical applications.

[0030] In summary, by introducing an empirical spatio-temporal variation function based on the wind propagation mechanism, the present invention effectively makes up for the deficiencies of existing wind power prediction methods in modeling small-scale spatio-temporal information and captures the interactions between wind turbines within a wind power cluster. And by improving the application process of the empirical spatio-temporal variation function, the computational complexity is reduced while ensuring the efficiency of the method. In addition, since the feature construction is based on the theoretical basis of the wind propagation effect, these features are more interpretable. At the same time, the present invention combines various system evaluation means to effectively screen the input features and eliminate redundant or weakly correlated variables. Compared with the traditional method of directly using large-scale meteorological factors without feature optimization, this method improves the prediction accuracy and model stability and provides a more accurate and practical wind power cluster output prediction scheme. BRIEF DESCRIPTION OF THE DRAWINGS

[0031] Figure 1 is a flowchart of the method according to an embodiment of the present invention; Figure 2 is a spatio-temporal interpolation flowchart. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0032] The present invention will be further clarified below in conjunction with specific embodiments. It should be understood that these embodiments are only used to illustrate the present invention and not to limit the scope of the present invention. After reading the present invention, those skilled in the art's various equivalent modifications of the present invention all fall within the scope defined by the appended claims of this application.

[0033] As Figure 1 described, the wind power cluster output prediction method with spatio-temporal feature coupling includes the following steps: Step 1: Collect data. In addition to the data of the wind power itself, the meteorological factors (such as wind speed, wind direction, temperature) that affect the wind power and the data of the wind turbine operating conditions (such as nacelle direction, blade angle) need to be sorted out and collected, and the time is aligned.

[0034] Step2: Spatio-temporal interpolation. The missing values in the target sequence (wind power data) and the feature data (meteorological factors and wind turbine operating conditions data, etc.) are classified and processed.

[0035] Step2.1: Preliminary screening and classification of missing values. For each missing value, first determine its neighboring observation points in the spatio-temporal domain. Based on the spatial distance, select two turbines with the closest physical distance and valid observations from the nearest 10 turbines as candidate points. At the same time, for the time dimension, define a reasonable time window and select the two time points closest to the missing moment before and after the target turbine with valid observations as candidates.

[0036] Step2.2: Apply corresponding interpolation strategies to handle missing values according to the situation.

[0037] (a) Bilinear interpolation: When the candidate points contain both valid spatial neighboring observations and data at two time stamps before and after, the weighted bilinear interpolation method is used for filling. This process takes into account the continuity in the time dimension and the strong correlation between adjacent observations of the same object, so a higher weight is given to the time axis; while for the spatial dimension, relatively lower but decreasing weights are assigned according to the physical distance.

[0038] Select neighboring valid observation points in both the time and space dimensions, and obtain the estimated value of the target point through weighted fusion. The calculation formula is as follows: x(·)=w s *( w s1 * x s1 + w s2 * x s2 )+ w t *( w t1 * x t1 + w t2 * x t2 ), where x(·) is the value used to fill the missing value. x s1 and x s2 respectively represent the observations of the spatial candidate points, x t1 and x t2 represent the observations of the time candidate points. w s and w t are the weighted coefficients for the overall spatial and time dimensions respectively, and satisfy w s +w t =1. To highlight the dominant role of time continuity in the change process, a higher weight is given to the time dimension (i.e., w s <w t ), thereby improving the rationality of the interpolation results in space and time. In addition, for the observations of the spatial candidate points, decreasing weights are assigned according to the physical distance. Use d1 and d2 to represent the distances to the two nearest spatial candidate points respectively, then the spatial weight w s1 = d1 / (d1+d2), w s2= d2 / ( d1+d2), thus ensuring that the spatial candidate points closer in distance have greater weights.

[0039] (b) Temporal linear interpolation: If more than 80% of the data in the spatial candidate points is missing, only the two valid timestamps before and after within the selected time window are relied on to perform linear interpolation. This method assumes that the time series data changes smoothly in a short period and is applicable to the situation where the data is continuous in time but insufficient in spatial information.

[0040] (c) Spatial nearest neighbor interpolation: When only spatial neighboring points are available, that is, no valid timestamp data can be obtained, then the two closest observations are selected as candidate points from the spatial dimension. Here, the linear interpolation method is adopted, but different from temporal linear interpolation, it assigns weights based on spatial distance, and the closer the observations are, the greater their weights. This method is applicable to the situation where the spatial distribution is relatively uniform and the temporal information is completely missing.

[0041] (d) Data deletion: For those cases where there are neither suitable spatial neighboring points nor neighboring observations on the time axis, it is considered that the information of these points is not sufficient to support reliable interpolation, so they are selected to be removed.

[0042] In addition, it is necessary to specifically explain the weight design principle. The most basic principle is that the sum of all weights during interpolation is 1, whether in the temporal dimension or the spatial dimension. And when implementing bilinear interpolation, the importance of the temporal dimension is particularly emphasized because it represents the results of continuous observations of the same individual. In contrast, although the spatial dimension can provide additional information, it depends on the similarity between different individuals, so its weight is set relatively low. At the same time, within the spatial dimension, the principle of "the closer the distance, the greater the correlation" is also followed, ensuring that the spatial weight gradually decreases as the distance increases.

[0043] Step3: Construct leading indicators. Take the deep learning model as the wind power cluster output prediction model and construct effective leading indicators as one of the model input features. The leading indicators contain information about the future wind power trend, so they are very helpful for predicting wind power. In addition, because when applying the wind propagation mechanism in the past, it was necessary to search for the traversal of the propagation time and locate the upstream wind turbines, the computational complexity was very high. However, in the embodiment of the present invention, by selecting an appropriate propagation time and fixing it, and then through the wind propagation mechanism, the same effective upstream wind turbines can be obtained with less computational complexity.

[0044] Step3.1: Determine the appropriate propagation time . By multiplying the propagation time by the average wind speed to obtain the propagation distance, this calculated distance is consistent with the actual distance between adjacent wind turbines in the wind power cluster. When the propagation time When it is too large, it may lead to the spatial granularity being too rough to accurately capture the changes within the wind field, and the time granularity needs to be increased to improve the accuracy. On the contrary, when the propagation time is too small, the spatial granularity is too fine, which may cause the time granularity to exceed the wind field boundary, and the time granularity needs to be reduced to ensure the validity of the data.

[0045] Step3.2: Determine the upstream wind turbines. Within the time window , calculate the empirical spatio-temporal variogram between all pairs of wind turbines . represents the target wind turbine, traverse all the remaining wind turbines, and then through the formula The selected wind turbine is used as the upstream wind turbine of the target wind turbine at this moment. Repeat this step until the upstream wind turbines at all moments are found.

[0046] Step 3.3: Judge the effectiveness of the upstream wind turbine. Judge whether the wind propagation effect of the upstream wind turbine on the downstream wind turbine is effective by whether the formula is less than 0.5, where represents the upstream wind turbine, represents the downstream wind turbine (which is also the target wind turbine for prediction). If , it is considered that the wind propagation effect between these two wind turbines is effective; on the contrary, if , it is considered that the wind propagation effect between these two wind turbines is ineffective.

[0047] Step3.4: Construct leading indicators. Use the historical data of the effective upstream wind turbines as leading indicators. Since the upstream and downstream relationships of the wind turbines in the wind power cluster are not static, and the relationships between the wind turbines will change with the change of the wind direction, it is necessary to note that the upstream wind turbines at time t are identified through the wind power cluster output data from time t - L to t - L + 1.

[0048] Step4: Feature selection.

[0049] Evaluate the relationship between wind power and various alternative features, and screen out the features useful for predicting the output of a wind power cluster. The alternative features are mainly divided into four categories, namely historical wind power data, meteorological factors (such as wind speed, wind direction, temperature, etc.), the operating conditions of wind turbines (such as nacelle direction, blade angle, etc.), and leading indicators (historical data of the upstream wind turbines corresponding to the target wind turbine). Three methods are used to examine the linear relationship, monotonic relationship, and complex non-linear relationship between wind power and alternative features respectively. These methods are Pearson correlation coefficient, Spearman correlation coefficient, and pre-experiment. The Pearson correlation coefficient is used to evaluate the strength and direction of the linear relationship between variables, while the Spearman correlation coefficient focuses on evaluating the monotonic association between variables, that is, the relationship where two variables are consistent in the change trend but not necessarily linear. In order to further reveal the non-linear dependence relationship that may not be fully captured by these traditional statistical methods, the present invention adopts a pre-experiment based on random forest. This method measures the importance of features by calculating the degree of average impurity reduction of each feature in all decision trees, so as to highlight the features that can most effectively split the data for prediction. Considering the results of these three methods comprehensively, screen out the effective features that are most important for the prediction of wind power cluster output, so as to provide a solid foundation for subsequent research.

[0050] Step5: Use a deep learning model for multi-step prediction. Construct a multi-input multi-output network structure, input the selected effective features into the deep learning model at one time, and directly output the prediction results of the wind power cluster output at multiple future time points from 15 minutes to 16 hours. In the initial stage, the internal parameters of the model are randomly initialized; subsequently, by minimizing the loss function (i.e., mean square error), use the backpropagation algorithm to continuously adjust the model parameters, and finally obtain a wind power output prediction model with optimized performance and higher prediction accuracy.

[0051] A wind power cluster output prediction system with spatio-temporal feature coupling, including the following modules: Data collection module; used to collect wind power data, meteorological factor data, and wind turbine operating condition data, and align the time of different data; the wind power data is the target sequence, and the meteorological factor data and wind turbine operating condition data are feature data; Spatio-temporal interpolation module; classify and process the missing values in the target sequence and feature data; Leading indicator construction module; by mining the spatio-temporal information between each wind turbine in the wind power cluster, identify the upstream and downstream relationships between each wind turbine and other wind turbines at each moment, and determine the upstream wind turbines by maximizing the wind propagation effect. Use all the first-order lag historical data of the upstream wind turbines as leading indicators, which are alternative features for predicting the output of downstream wind turbines; Feature selection module: It evaluates the relationship between wind power and various alternative features, and screens out the features useful for predicting the output of a wind power cluster. The alternative features include four categories, namely historical wind power data, meteorological factors, wind turbine operating conditions, and leading indicators. Three methods, namely Pearson correlation coefficient, Spearman correlation coefficient, and pre-experiment based on random forest, are used to examine the linear relationship, monotonic relationship, and complex non-linear relationship between wind power and alternative features respectively. The union of the features with high linear correlation, strong monotonic relationship, and high random forest importance score is taken as the input feature set of the prediction model, and the features that perform poorly in all three evaluation criteria are discarded, so as to ensure the effectiveness of the features input into the prediction model. Prediction module: It uses a deep learning model for multi-step prediction. The selected effective features are input into a multi-input multi-output deep learning model, and the prediction results of the output of the wind power cluster at multiple future time points within a set time period are obtained.

[0052] To verify the effectiveness of the present invention, two groups of predictions were made on 134 wind turbines using five state-of-the-art deep learning models. One group only used historical wind power data, while the other group used the effective features obtained through screening. The results show that the input of effective features including leading indicators significantly improves the model performance, and is more superior in terms of the Mean Absolute Percentage Error (MAPE) index, indicating that the constructed leading indicators and feature selection module have a significant effect on improving the prediction accuracy of wind power generation. Through the comparison of experimental results, it is further confirmed that the introduction of leading indicators and feature selection can effectively improve the performance of the prediction model, providing strong support for the accurate prediction of wind power generation. This finding not only verifies the effectiveness of the method of the present invention, but also provides an important reference for future related research.

[0053] Obviously, those skilled in the art should understand that each step of the above-described wind power cluster output prediction method with spatio-temporal feature coupling or each module of the wind power cluster output prediction system with spatio-temporal feature coupling can be implemented by a general-purpose computing device. They can be concentrated on a single computing device or distributed on a network composed of multiple computing devices. Optionally, they can be implemented by program codes executable by the computing device, so that they can be stored in a storage device and executed by the computing device. And in some cases, the steps shown or described can be executed in a different order from here, or they can be made into individual integrated circuit modules respectively, or multiple modules or steps among them can be made into a single integrated circuit module to implement. In this way, the embodiments of the present invention are not limited to any specific combination of hardware and software.

Claims

1. A method for predicting the output of a wind power cluster with spatio-temporal feature coupling, characterized in that It includes the following steps: Step 1: Collect data; the data includes wind power data, meteorological factor data, and wind turbine operation status data, and align the time of different data; the wind power data is the target sequence, and the meteorological factor data and wind turbine operation status data are feature data; Step 2: Spatiotemporal interpolation; classify and process the missing values in the target sequence and feature data; Step 3: Construct leading indicators; use the deep learning model as the wind power cluster output prediction model, and construct leading indicators as one of the input features of the prediction model; by mining the spatiotemporal information between each wind turbine in the wind power cluster, identify the upstream and downstream relationships between each wind turbine and other wind turbines at each moment, and determine the upstream wind turbines by maximizing the wind propagation effect. Use all the first-order lag historical data of the upstream wind turbines as the leading indicators, which are the alternative features for predicting the output of downstream wind turbines; Step 4: Feature selection; evaluate the correlation and prediction ability between wind power and various alternative features, and screen out the input feature set of the wind power cluster output prediction model; The alternative features include four categories, namely historical wind power data, meteorological factors, wind turbine operation status, and leading indicators; use three methods, namely Pearson correlation coefficient, Spearman correlation coefficient, and pre-experiment based on random forest, to examine the linear relationship, monotonic relationship, and complex non-linear relationship between wind power and alternative features respectively. Take the union of the features with linear correlation reaching the set threshold, monotonic relationship reaching the set threshold, and random forest importance score reaching the set threshold as the input feature set of the wind power cluster output prediction model; Step 5: Use the deep learning model for multi-step prediction; input the selected effective features into the multi-input multi-output deep learning model, and output the wind power cluster output prediction results at multiple future time points within a set time period.

2. The wind power cluster output prediction method based on spatio-temporal feature coupling according to claim 1, wherein, In Step 5, at the initial stage of using the deep learning model for prediction, the internal parameters of the model are randomly initialized; subsequently, by minimizing the loss function, the model parameters are continuously adjusted using the backpropagation algorithm, and finally the wind power output prediction model is obtained.

3. The wind power cluster output prediction method with spatio-temporal feature coupling according to claim 1, wherein In Step 1, the meteorological factors include wind speed, wind direction, and temperature; the wind turbine operation status includes nacelle direction and blade angle.

4. The wind power cluster output prediction method based on spatio-temporal feature coupling according to claim 1, wherein In Step 2, for the missing values in the target sequence and feature data, classify and process them, specifically including the following content: Step 2.1: Preliminary screening and classification of missing values; for each missing value, first determine its neighboring observation points in the spatiotemporal domain; based on the spatial distance, select two wind turbines with the closest physical distance and valid observations as candidates; at the same time, for the time dimension, define a time window and select the two time points closest to the missing moment before and after the target wind turbine and with valid observations as candidates; Step 2.2: Apply corresponding interpolation strategies to handle the missing values according to the situation; (a) Bilinear interpolation: When the candidate point contains both valid spatial neighboring observations and data at two timestamps (before and after), the weighted bilinear interpolation method is used for filling; neighboring valid observation points are selected in both the time and space dimensions, and the estimated value of the target point is obtained through weighted fusion. The calculation formula for the estimated value of the target point is as follows: x(·)=w s *( w s1 * x s1 + w s2 * x s2 )+ w t *( w t1 * x t1 + w t2 * x t2 ) where x(·) is the value used to fill in the missing value; x s1 and x s2 represent the observed values of the spatial candidate points, x t1 and x t2 represent the observed values of the temporal candidate points; w s and w t are the weighting coefficients for the overall spatial and temporal dimensions respectively, and satisfy w s +w t = 1, w s <w t , for the observed values of the spatial candidate points, decreasing weights are assigned according to the physical distance; (b) Temporal linear interpolation: If more than a set proportion of the data in the spatial candidate points is missing, linear interpolation is performed relying only on the data at two valid timestamps (before and after) within the selected time window. (c) Spatial nearest neighbor interpolation: When only spatial neighboring points are available, the two closest observed values are selected from the spatial dimension as candidate points. (d) Data deletion: Remove the data that has neither spatial neighboring points nor neighboring observations on the time axis.

5. The method for predicting the output of a wind power cluster with spatio-temporal feature coupling according to claim 4, wherein In Step 2.2, the weight design principle is that the sum of all weights during interpolation is 1, regardless of whether it is in the time dimension or the space dimension.

6. The wind power cluster output prediction method based on spatio-temporal feature coupling according to claim 4, wherein In the said Step 2.2, for the observed values of the spatial candidate points, decreasing weights are assigned according to the physical distance; using d1 and d2 to represent the distances to the two nearest spatial candidate points respectively, then the spatial weight w s1 = d1 / ( d1+d2), w s2 = d2 / ( d1+d2).

7. The wind power cluster output prediction method based on spatio-temporal feature coupling according to claim 1, wherein In Step 3, the wind propagation mechanism is defined as follows: Suppose a gust of wind blows from wind turbine s1 to another wind turbine s2. During the propagation time h minutes, the wind speed of wind turbine s1 at time t is equal to the wind speed of wind turbine s2 at time t + h. The wind propagation effect, that is, the empirical spatio-temporal variogram g(s1, s2, h), is defined as: ， where x(s, t) represents the wind speed of wind turbine s at time t, and L represents the length of the time window; when the wind reaches wind turbine s2 from wind turbine s1 after a propagation time of h, g(s1, s2, h) = 0. Identify the most effective upstream wind turbines corresponding to each wind turbine at each moment through the wind propagation mechanism, and construct leading indicators using the historical data of the upstream wind turbines.

8. The wind power cluster output prediction method based on spatio-temporal feature coupling according to claim 1, wherein, Step 3 specifically includes the following steps: Step 3.1: Determine the propagation time , and obtain the propagation distance by multiplying the propagation time by the average wind speed. The propagation distance is consistent with the actual distance between adjacent wind turbines in the wind power cluster; Step 3.2: Determine the upstream wind turbines; the process of determining the upstream wind turbines of a single target wind turbine is as follows: within the time window L, calculate the empirical spatio-temporal variogram between all pairs of wind turbines ; s2 represents the target wind turbine, s1 represents all other wind turbines, and then use the formula to select the wind turbines as the upstream wind turbines of the target wind turbine at the current moment, thereby finding the upstream wind turbines of a single target wind turbine; by traversing all wind turbines in the cluster as target wind turbines and repeating the above process of determining the upstream wind turbines of a single target wind turbine, identify the upstream wind turbines of each wind turbine; on this basis, combined with the method of rolling time window, further identify the upstream wind turbines at different times, so as to realize the identification of upstream wind turbines in the whole-time wind power cluster; Step 3.3: Construct leading indicators; Use the historical data of the upstream wind turbines as leading indicators. Since the upstream and downstream relationships between the wind turbines in the wind power cluster change with the change of the wind direction, the upstream wind turbines at time t are identified through the output data of the wind power cluster from time t - L to t - L + 1.

9. A wind power cluster output prediction system with spatio-temporal feature coupling, characterized in that It includes the following modules: Data collection module; Used to collect wind power data, meteorological factor data, and wind turbine operation status data, and align the time of different data; the wind power data is the target sequence, and the meteorological factor data and wind turbine operation status data are feature data. Spatio-temporal interpolation module; Classify and process the missing values in the target sequence and feature data. Leading indicator construction module; By mining the spatio-temporal information between each wind turbine in the wind power cluster, identify the upstream and downstream relationships between each wind turbine and other wind turbines at each moment, and determine the upstream wind turbines by maximizing the wind propagation effect. Use all the first-order lag historical data of the upstream wind turbines as leading indicators, which are used as alternative features for predicting the output of the downstream wind turbines. Feature selection module; Evaluate the correlation and predictive ability between wind power and various alternative features, and screen out the input feature set for the wind power cluster output prediction model. The alternative features include four categories, namely historical wind power data, meteorological factors, wind turbine operating conditions, and leading indicators. Three methods, namely Pearson correlation coefficient, Spearman correlation coefficient, and pre-experiment based on random forest, are used to examine the linear relationship, monotonic relationship, and complex non-linear relationship between wind power and alternative features respectively. The union of the features with linear correlation reaching the set threshold, monotonic relationship reaching the set threshold, and random forest importance score reaching the set threshold is taken as the input feature set of the wind power cluster output prediction model. Prediction module; Use a deep learning model for multi-step prediction; input the selected effective features into a multi-input multi-output deep learning model, and output the wind power cluster output prediction results at multiple future time points within a set time period.

10. A computer device, characterized in that: The computer device includes a memory, a processor, and a computer program stored on the memory and executable on the processor. When the processor executes the above computer program, it implements the spatio-temporal feature coupling-based wind power cluster output prediction method as described in any one of claims 1-8.

Images