Ultra-short-term wind farm power multi-step interval prediction modular method and device thereof

Abstract

The modular method and device for ultra-short-term wind farm power multi-step interval prediction comprises the following steps: obtaining the geographical location information of the wind farm, and collecting and storing the historical operation data of the wind farm at a certain sampling interval; based on the geographical information and the historical operation data of the wind farm, the selection of the characteristic wind turbine of the wind farm is comprehensively realized, and the multi-element characteristic wind condition of the wind farm is obtained on this basis; based on the selection result of the characteristic wind condition of the wind farm, the ultra-short-term multi-step prediction of the characteristic wind turbine wind condition is realized through different time finite difference direct time sequence modeling strategies, and a multi-step interval prediction model is established; the finite difference operation domain is defined, and the convex partition of the finite difference operation space is realized; the global time finite difference-machine learning wind farm aggregate power generation characteristic modeling is realized; based on the established characteristic wind condition prediction model and the wind farm aggregate power generation characteristic model, the rolling prediction mechanism is used to realize the ultra-short-term multi-step interval prediction of the wind farm.

Description

Technical Field

[0001] This invention relates to a method for predicting the output power of a wind farm, specifically to a modular method and equipment for multi-step interval prediction of ultra-short-term wind farm power. Background Technology

[0002] Compared to traditional non-renewable energy sources, renewable energy is widely distributed globally and has abundant reserves. Wind energy is a clean, renewable, and green energy source with the potential for long-term reliance. Among many new energy sources, wind power has advantages such as low generation costs, abundant resources, and clean and efficient operation. However, the uncertainty of wind is one of the main problems in wind power generation. Due to the uncertainty of weather conditions, wind power has poor predictability and low prediction reliability. The negative impacts on the power grid after large-scale wind power grid connection, such as wind curtailment, voltage and frequency fluctuations, remain pressing issues that need to be addressed. The instability of wind power often restricts grid dispatch after grid connection, preventing the timely and effective utilization of generated wind power. To further improve wind energy utilization and promote wind energy consumption, researching high-precision wind power prediction models is of great significance. By obtaining high-precision wind power prediction values ​​for future moments, power sectors can formulate reasonable dispatch plans, arrange the grid connection capacity of various wind farms in real time, and improve the grid's ability to accept unstable wind power generation. This provides fundamental support for grid peak and frequency regulation, safe and stable operation, and economic management.

[0003] As wind power installed capacity continues to increase, the demand for accurate wind power forecasts is also growing. Wind power forecasting refers to obtaining the output power of a single wind turbine or the entire wind farm within a certain future time period. However, due to the randomness and fluctuation of wind, wind power is inevitably affected by many factors such as wind shear, turbulence intensity, temperature, pressure, and air density. Accurate and effective wind power forecasting methods are currently a research challenge in wind power systems.

[0004] In wind power forecasting research, based on different time scales, wind power forecasting techniques can be divided into medium-to-long-term time-scale forecasting, short-term time-scale forecasting, and ultra-short-term time-scale forecasting. Medium-to-long-term forecasts cover the next few months to several years, providing assistance and reference for wind farm construction planning, layout design, and operational scheme verification. Short-term forecasts cover the next few hours to several days, helping grid technicians develop economically reasonable dispatch plans to ensure timely grid connection for wind power. Ultra-short-term forecasts, measured in minutes, predict the next few minutes or hours, primarily used for wind farm operation control, real-time power system settlement, and the formulation of intraday and real-time power generation plans.

[0005] The ultimate goal of wind power forecasting is to obtain predicted wind power values. Depending on the different objects being forecasted, wind power forecasting can be divided into direct forecasting and indirect forecasting. The former directly obtains predicted wind power values ​​through historical wind power data and forecasting models, while the latter involves two steps: wind resource forecasting and wind power generation characteristic modeling. This latter approach better reflects the causal relationship of wind power generation and helps to flexibly control the accuracy of forecasts.

[0006] In recent years, time series analysis, support vector machines (SVMs), and neural networks have been widely applied to ultra-short-term wind resource forecasting and wind power generation characteristic modeling. To achieve high-precision wind power forecasting, selecting better data inputs based on feature engineering is crucial. Most research focuses on the analysis of operational data, lacking comprehensive consideration of wind farm geographical information. It is worth noting that in wind power generation characteristic modeling research, autoregressive models, artificial neural networks, and SVMs are static modeling methods. These models cannot automatically extract deep features from the data, significantly limiting their accuracy. Deep learning algorithms in machine learning, on the other hand, not only possess flexible network structures but also strong nonlinear fitting capabilities, effectively enabling dynamic modeling of time series data. They have been widely used in ultra-short-term wind power forecasting in recent years. Furthermore, current literature on the time delay between wind resources and wind farm output is lacking, requiring further research. To refine the characterization of forecast uncertainty, many scholars have begun to study wind power probability interval models that include forecast intervals or conditional probability distributions. Commonly used methods include quantile regression, upper and lower limit estimation, Gaussian mixture model, and kernel density estimation. Reliable interval prediction can further meet the needs of power system risk assessment and decision-making, and plays an important role in the safe and stable operation of the power system.

[0007] Due to the influence of factors such as the spatiotemporal dispersion of wind, wind turbine units, and their output power, ultra-short-term wind power prediction remains a significant challenge. In particular, if the geographical and wind conditions of a wind farm are highly unstable, the operation of the wind turbines exhibits a high degree of chaotic complexity, making ultra-short-term prediction of their combined output power extremely difficult. Therefore, building upon existing research, it is of great importance to study how to establish a high-precision ultra-short-term wind power prediction model. Summary of the Invention

[0008] To address the shortcomings of existing technologies, this invention proposes a modular method and device for multi-step interval prediction of ultra-short-term wind farm power, employing stepwise prediction. First, considering the spatiotemporal dispersion of wind conditions, a multi-characteristic wind condition of the wind farm is selected based on wind farm operation data and geographical information to reduce model complexity. Then, based on the time dependence of wind resource prediction, an ultra-short-term multi-step interval prediction model for characteristic wind turbine wind conditions is established using different TFD direct time series modeling strategies. The prediction accuracy of characteristic wind conditions is ensured by comparing the effects of different modeling strategies. Second, given the complex wind conditions and time delays between power outputs in wind farms, a finite difference operating domain is constructed. A global TFD-ML modeling method is proposed to model the comprehensive power generation characteristics of the wind farm between the dispersed multiple wind conditions and the total wind farm power. Based on this, a flexible ultra-short-term wind power prediction scheme for wind farms can be established. Finally, a multi-step interval prediction model for wind power is established using non-parametric and semi-parametric interval construction algorithms, and its technical solution is as follows:

[0009] A modular method for multi-step interval prediction of ultra-short-term wind farm power, characterized by the following steps:

[0010] Step 1: Obtain the geographical location information of the wind farm and collect and store the historical operation data of the wind farm at certain sampling intervals.

[0011] Step 2: Based on the geographical information and historical operation data of the wind farm, comprehensively select wind turbines with wind farm characteristics, and obtain the multi-characteristic wind conditions of the wind farm.

[0012] Step 3: Based on the wind condition selection results of the wind farm characteristics, the ultra-short-term multi-step prediction of the wind condition of the characteristic wind turbine is realized through different temporal finite difference (TFD) direct time series modeling strategies. On this basis, a multi-step interval prediction model is established.

[0013] Step 4: Define the finite difference operating domain and implement the convex partitioning of the finite difference computation space. Based on this, implement global temporal finite difference (TFD)-machine learning (ML) modeling of the lumped generation characteristics of wind farms.

[0014] Step 5: Based on the established characteristic wind condition prediction model and wind farm lumped generation characteristic model, and based on the rolling prediction mechanism, realize ultra-short-term multi-step interval prediction of wind farm.

[0015] The present invention also discloses a non-volatile storage medium, characterized in that the non-volatile storage medium includes a stored program, wherein the program, when running, controls the device where the non-volatile storage medium is located to execute the above-described method.

[0016] The present invention also discloses an electronic device, characterized in that it comprises a processor and a memory; the memory stores computer-readable instructions, and the processor is used to execute the computer-readable instructions, wherein the computer-readable instructions execute the method described above.

[0017] Beneficial effects

[0018] This invention includes wind farm data acquisition and storage, characteristic wind condition extraction, multi-dimensional characteristic wind condition interval prediction model and wind farm lumped output characteristic model. Through step-by-step prediction, it ultimately achieves multi-step interval prediction of ultra-short-term wind power.

[0019] This invention proposes a multi-step interval prediction method for ultra-short-term wind farm power based on characteristic wind condition selection and time-domain finite-difference dynamic modeling, which can effectively improve the accuracy of ultra-short-term wind farm prediction. While ensuring the accuracy of wind power prediction, it further quantifies the uncertainty of the model, and the established multi-step interval prediction model for wind power has high reliability and strong application value, which is conducive to improving the grid-connected capacity of wind power and the stability of wind power grid-connected system operation. Attached Figure Description

[0020] Figure 1 This is a flowchart illustrating the implementation of a modular method for multi-step interval prediction of ultra-short-term wind farm power according to the present invention.

[0021] Figure 2 This is a curve comparing the final prediction results of the present invention. Detailed Implementation

[0022] Given the challenges posed by the time delay and spatiotemporal dispersion of wind turbine wind conditions and output to ultra-short-term wind power prediction, the purpose of this invention is to provide an ultra-short-term wind farm power range prediction system and its modular method.

[0023] A method for predicting the output power of a wind farm, specifically involving a modular method for multi-step interval prediction of ultra-short-term wind farm power, characterized by the following steps:

[0024] Step 1: Obtain the geographical location information of the wind farm and collect and store the historical operation data of the wind farm at certain sampling intervals.

[0025] The geographical location information of the wind farm includes the longitude, latitude, and altitude of all wind turbines within the wind farm; the sampling interval must meet the requirements of ultra-short-term wind power prediction, and the sampling time interval defined for ultra-short-term power prediction is 1 to 15 minutes; the historical operating data of the wind farm includes historical data on the active power, wind speed, and wind direction of all wind turbines in the wind farm, characterized in that the selected wind power data must undergo abnormal data cleaning processing in order to provide a data sample basis for subsequent ultra-short-term wind power prediction.

[0026] Step 2: Based on the geographical information and historical operation data of the wind farm, comprehensively select wind turbines with wind farm characteristics, and obtain the multi-characteristic wind conditions of the wind farm.

[0027] The wind turbine selection method based on wind farm characteristics includes two parts: external correlation analysis and internal correlation analysis.

[0028] First, based on historical operational data, an external correlation analysis of the wind farm is conducted to quantify the impact of wind speed and direction of all wind turbines on the wind farm's power output. Then, by determining a suitable feature selection algorithm, a feature selection model is established to analyze the nonlinear relationship between input elements and target output. Using the feature weights returned by the feature selection model, the potential wind turbine weights (WT) of the wind farm are preliminarily determined. h .

[0029] Secondly, combining geographical information of wind farms, we discuss the correlation between wind conditions within wind farms through internal correlation analysis to avoid information redundancy leading to missing wind condition information. We quantify the geographical information between wind turbine units, determine appropriate nonlinear correlation analysis algorithms to measure the correlation between two random variables, and combine geographical information and nonlinear correlation calculation results to exclude candidate wind turbines (WT) from wind farms. h The redundant fans in the system are used to determine the final characteristic fan WT.

[0030] Ultimately, by using the wind turbine selection method based on the characteristics of the wind farm, the wind conditions of the selected characteristic wind turbines represent the multi-dimensional characteristic wind conditions of the wind farm.

[0031] The aforementioned external correlation analysis is characterized by simultaneously considering the impact of both wind speed and wind direction on wind farm power. Therefore, the relationship between the two is discussed in two cases: a) If the correlation between wind speed and wind direction is strong, a corresponding dimensionality reduction algorithm is determined to reduce the dimensionality of the two-dimensional data composed of wind speed and wind direction for each wind turbine. Dimensionality reduction data D dim The higher the level of feature information contained in the original variables, the better the dimensionality reduction effect of the high-dimensional data. The dimensionality-reduced data D... dim Normalized wind farm power is used as the input and output of the feature selection model, respectively. Wind turbines with larger feature weights are selected as candidate wind turbines (WT). hb) If the correlation between wind speed and wind direction is weak, and the data after dimensionality reduction does not contain much feature information of the original variables, it indicates that wind speed and wind direction are two relatively independent variables. In this case, two feature selection models are established to calculate the impact of wind turbine wind speed and wind direction on wind farm power separately. Wind turbines with larger feature weights in both models are selected as candidate wind turbines (WT). h .

[0032] The internal correlation analysis is characterized by quantifying the geographical location information within the wind farm and completing the analysis of the nonlinear correlation of wind conditions between wind turbine units, thereby eliminating redundant wind turbines. The quantification of geographical location information is characterized by normalizing the longitude, latitude, and altitude information of each operating unit within the wind farm and then calculating the Euclidean distance between each pair of units. The wind turbine unit wind condition nonlinear correlation analysis method is characterized by determining the nonlinear correlation analysis algorithm and performing wind condition correlation analysis in two cases based on the different strengths of the correlation between wind speed and wind direction, specifically including the following: a) If the correlation between wind speed and wind direction is strong, calculate the dimensionality-reduced data D of each wind turbine unit. dim a) If the correlation between wind speed and wind direction is weak, calculate the nonlinear correlation between the original wind speeds and the original wind directions of each wind turbine. The method for eliminating redundant wind turbines is characterized by making a comprehensive judgment based on the normalized Euclidean distance between wind turbines and the results of wind condition correlation analysis. If the Euclidean distance between two wind turbines is small and the wind condition correlation is high, then they are considered to be geographically close and have similar wind condition information, and are thus determined to be redundant wind turbines. Based on the feature selection model results in the external correlation analysis, wind turbines with smaller feature weights are eliminated, thereby achieving the elimination of redundant wind turbines.

[0033] The wind turbine selection method for wind farm characteristics is characterized by the following steps: Step 1. Selecting candidate wind turbines (WT) through external correlation analysis. h Step 2. Quantify the geographical location information within the wind farm and calculate the nonlinear correlation of wind conditions for each wind turbine. Step 3. Analyze the wind condition correlation of each wind turbine. Based on the redundancy determination principle, remove f1 redundant wind turbines with high wind condition correlation and small Euclidean distance. Step 4. Reselect f2 wind turbines with larger feature weights from the remaining turbines and add them to the WT (Wind Wrapper). h Step 5. Repeat Steps 3 and 4 to analyze the wind turbine's nonlinear correlation value with wind conditions again, until WT. h There are no redundant fans. The final multi-characteristic fan WT = {WT1, WT2, ..., WT} is determined. Q}, where Q represents the number of characteristic fans.

[0034] Step 3: Based on the wind condition selection results of the wind farm characteristics, ultra-short-term multi-step prediction of the wind conditions of characteristic wind turbines is realized through different temporal finite difference (TFD) direct time series modeling strategies. On this basis, a multi-step interval prediction model is established.

[0035] The TFD direct time series modeling strategy includes the following: based on the TFD modeling concept of univariate time series, it simultaneously employs both multiple inputs single output (MISO)-TFD modeling and multiple inputs multiple outputs (MIMO)-TFD modeling strategies to ensure the reliability and flexibility of wind condition forecasting. The univariate time series TFD modeling concept is characterized by the wind condition Y from time t+1 to time t+Δt. wc (t+1, t+Δt), has a certain temporal correlation with the wind conditions before time t+1. Let the wind conditions before time t+1 be a finite difference wind condition vector И(t)=[Y wc (t),Y wc (t-1),...,Y wc (tn x If ), then the mathematical relationship between them can be described as: Y wc (t+1,t+Δt)=f(И(t))+e c (t). Wherein, Y wc (t+1,t+Δt)=[Y wc (t+1),Y wc (t+2),...,Y wc (t+n y )], Y wc For a wind farm, the wind speed or wind direction time series is n. x and n y These represent the input and output orders of the model, respectively, determined by selecting the appropriate model order determination algorithm; f(·) represents a finite number of difference operations; e c The signal is noise. The two different modeling strategies are characterized by determining the corresponding advanced machine learning neural network algorithm to establish a feature-based wind condition prediction model. For the MISO-TFD modeling strategy, the model input is multi-dimensional, and the output is one-dimensional, by using multiple inputs (n... x ≥1) and a single output (n) y =1) Establish a characteristic wind condition prediction model. Through a rolling prediction mechanism, the predicted values ​​are used as input to the characteristic wind condition model to achieve multi-step advanced prediction of characteristic wind conditions; for the MIMO-TFD modeling strategy, the model input is multi-dimensional, and the output is also multi-dimensional, by using multiple inputs (nx ≥1) and a single output (n) y ≥2) Establish a feature-based wind condition prediction model. Based on machine learning neural network algorithms, construct the mapping relationship between multi-dimensional inputs and multi-dimensional outputs to realize feature-based wind condition prediction. The multi-dimensional output of the model is the required advanced multi-step prediction result.

[0036] The characteristic wind condition multi-step interval prediction model is characterized by determining the corresponding interval construction algorithm based on the multi-step wind condition prediction results, and simultaneously using non-parametric and semi-parametric interval construction algorithms to establish the multi-step interval prediction model, effectively ensuring the reliability of the uncertain interval model.

[0037] Step 4: Define the finite difference operating domain and implement the convex partitioning of the finite difference computation space. Based on this, implement global temporal finite difference (TFD)-machine learning (ML) modeling of the lumped generation characteristics of wind farms.

[0038] The finite difference operating domain is constituted by the space spanned by the TFD hybrid eigenvector φ(t). The TFD hybrid eigenvector φ(t) is characterized in that φ(t) includes the wind direction and speed of all characteristic wind conditions, and encompasses the time delay relationship between these wind conditions and the wind farm's output power, specifically φ(t) = [P(t-1),...,P(tn)]. a ),V1(t),V1(t-1),...,V1(tm a ),...,V Q (t),V Q (t-1),...,V Q (tm a ),W1(t),W1(t-1),...,W1(tm a ),...,W Q (t),W Q (t-1),...,W Q (tm a ]], where P is the output power of the wind farm, and V and W are the characteristic wind turbine wind speed and wind direction; n a and m a The model order is determined by selecting the appropriate model order determination algorithm.

[0039] The convex partitioning of the finite difference operation space is characterized by considering the periodic wind direction data present in φ(t), and redefining a new and suitable similarity metric d(φ) according to the traditional partition-based clustering algorithm process. norm,i (t),φ norm,j (t) and clustering performance evaluation index O SSEThis completes the classification of the data.

[0040] The similarity metric d(φ) norm,i (t),φ norm,j (t)), characterized in that its calculation method includes the following steps: Step 1. Let the wind direction component variable Φ2(t) = {w i |i=1,2,...,N}, where N is the number of samples, and for any two wind direction samples w i and w j Calculate the angle of wind direction change W wd Step 2. Normalize the power and wind speed in the TFD mixed feature vector φ(t) to the interval [-1,1] to form the variable Φ1(t)={x i |i=1,2,...,N}, where N is the number of samples. Define the normalized TFD mixture feature vector φ. norm (t)=[Φ1(t),Φ2(t)], for any two samples, the new similarity metric is calculated as follows:

[0041]

[0042] Where φ norm,i (t)=[x i1 ,x i2 ,...,x ih ,w i1 ,w i2 ,...,w iM ], M is the dimension of the wind direction data w; φ norm,j (t)=[x j1 ,x j2 ,...,x jh ,w j1 ,w j2 ,...,w jM h is the dimension of variable x, i,j=1,2,...,N.

[0043] The clustering effect evaluation index O SSE Its characteristic is that its specific calculation formula is as follows:

[0044]

[0045] Where D s The subset of data obtained after clustering; C is the number of clusters; φ norm,s (t) represents the cluster center, φ norm,R (t) represents the subset D s The remaining sample points after removing the cluster centers;

[0046] The global TFD-ML wind farm lumped generation characteristic modeling is characterized by flexibly selecting different modeling schemes according to the varying complexity of wind farm operation. Specifically, it includes the following steps: Step 1. Based on the finite difference operation partitioning results, select samples in each sub-operational domain. Step 2. Select machine learning algorithms and cross-validation mechanisms to establish a unified model and sub-operational domain models. The unified model constructs the input-output mapping relationship across the entire finite difference domain, while the sub-operational domain models establish input-output mapping models separately in each sub-operational domain. Step 3. Analyze the selected wind farm operation complexity. If the accuracy of the two models is similar, it indicates that the machine learning algorithm has sufficient global learning capability, and the unified model is sufficient; if the latter has higher accuracy, it indicates that the nonlinearity of the modeling process is too complex, exceeding the learning capability of the machine learning algorithm, and it is recommended to use the sub-operational model. Step 4. Select the model with the best performance as the final wind farm lumped generation characteristic modeling scheme to improve modeling flexibility and effectively address the complex nonlinearity and uncertainty of wind farms.

[0047] The unified model and the sub-operational domain model are characterized by using finite difference regression vectors. The wind farm output power P(t) is used as the model input and modeled as the model output.

[0048] The finite difference regression vector Its key feature is that, considering the most direct energy conversion relationship between wind speed and wind power, a time-finite difference regression vector is defined based on the wind speed of all characteristic wind conditions and its time delay relationship with wind power.

[0049] Step 5: Based on the established characteristic wind condition prediction model and wind farm lumped generation characteristic model, and based on the rolling prediction mechanism, realize ultra-short-term multi-step interval prediction of wind farm.

[0050] The method includes the following steps: Step 1. Generate a TFD hybrid feature vector φ(t) based on wind speed and direction prediction information. Step 2. Determine the operating domain of the convex partitioning result of the finite difference operation space and select the corresponding prediction model, which can be a unified model or a sub-operating domain model. Step 3. Input the ultra-short-term wind speed interval prediction result into the global time finite difference (TFD)-machine learning (ML) model to obtain the ultra-short-term multi-step prediction result of wind farm output power. Step 4. Based on the multi-step prediction result of wind farm wind power, determine the prediction interval construction algorithm, and simultaneously use non-parametric and semi-parametric interval construction algorithms to establish an ultra-short-term wind power multi-step interval prediction model to effectively ensure the reliability of the uncertain interval model. Step 5. Evaluate and monitor the prediction accuracy online. If the prediction accuracy continues to be lower than the set requirements, the TFD-ML model needs to be updated.

[0051] Example

[0052] The following, in conjunction with the accompanying drawings and preferred embodiments of the present invention, further illustrates the technical means employed by the present invention to achieve its intended purpose. The present invention selects a wind farm in southwest my country, which has 24 wind turbines with an installed capacity of 48MW. The sampling time interval for historical operating data of wind speed, wind direction, and wind power of the wind farm is selected as 1 minute to meet the ultra-short-term prediction requirements proposed in this invention. After data cleaning and processing, these data serve as training and testing samples for the ultra-short-term wind farm power range prediction system proposed in this invention.

[0053] Figure 1 This is a flowchart illustrating the implementation of an ultra-short-term wind farm power range prediction system and modular method. The system includes a characteristic wind condition extraction module, a characteristic wind condition prediction module, a wind farm lumped output characteristic modeling module, and a wind power multi-step range prediction module, ultimately achieving ultra-short-term wind power multi-step range prediction. Specifically, it includes the following steps:

[0054] Step 1: Obtain the geographical location information of the wind farm and collect and store historical operational data of the wind farm at certain sampling intervals. Specifically, this includes: obtaining geographical location information such as the longitude, latitude, and altitude of the wind farm; obtaining historical operational data on wind direction, wind speed, and wind power of the wind farm at certain sampling intervals, which must meet the requirements of ultra-short-term wind power prediction. The sampling time interval defined for ultra-short-term power prediction is 1 to 15 minutes; after data cleaning, this data is used as the basic modeling dataset.

[0055] Step Two: Based on the geographical information and historical operation data of the wind farm, a comprehensive selection of wind turbines with distinctive characteristics is achieved, and multi-dimensional wind conditions of the wind farm are obtained on this basis. The method for selecting wind turbines with distinctive characteristics includes two parts: external correlation analysis and internal correlation analysis. Specifically, it includes: First, based on the historical operation data of the wind farm, the candidate wind turbines (WT) are initially determined according to the external correlation analysis. h First, the appropriate feature selection algorithm is determined. In this case, neighborhood components analysis (NCA) is preferred to determine the complex relationship between wind speed, wind direction, and power. NCA is a non-parametric method for feature selection of nonlinear input-output relationships, and the returned feature weights can measure the degree of influence of the input on the output. Since both wind speed and wind direction are considered simultaneously, their relationship is discussed first:

[0056] (1) If there is a strong correlation between wind speed and wind direction in wind turbines, a suitable dimensionality reduction algorithm should be determined for data dimensionality reduction. In this case, principal component analysis (PCA) is used to reduce the dimensionality of the original data. The dimensionality-reduced data D... dim Representing the original wind speed and direction, the contribution rate reflects the proportion of original information. (D) dim Normalized wind farm power was used as the input and output of the NCA model, respectively, and wind turbines with larger feature weights were selected as candidate wind turbines (WT). h .

[0057] (2) If the correlation between wind speed and wind direction of the wind turbine is low, the contribution rate after PCA dimensionality reduction is not high. This indicates that wind speed and wind direction are two relatively independent variables, and the influence of wind speed and wind direction on wind farm power can be calculated separately. Therefore, two NCA models are established. The inputs of the two NCA models are the wind speed and wind direction of each wind turbine, and the outputs are the normalized wind farm power. Wind turbines with larger feature weights in both models are selected as candidate wind turbines WT. h .

[0058] Secondly, combining the geographical information of wind farms, this paper discusses the correlation between wind conditions within wind farms through internal correlation analysis to avoid information redundancy leading to missing wind condition information. First, a suitable nonlinear correlation analysis algorithm is determined. In this case, the maximum information coefficient (MIC) is used to analyze the nonlinear correlation between wind turbine wind speeds and directions. The MIC value ranges from 0 to 1; the closer it is to 1, the stronger the correlation, and vice versa.

[0059] The selection method for characteristic fans is as follows:

[0060] Step 1: Determine candidate wind turbines (WT) based on external correlation analysis. h .

[0061] Step 2. Quantify the geographical information of the wind farm. After normalizing longitude, latitude, and altitude, calculate the Euclidean distance between each wind turbine. Based on the different correlations between wind speed and wind direction, select the dimensionality-reduced data D for each wind turbine to calculate. dim The MIC, or calculate the MIC between the original wind speeds and wind directions of each wind turbine.

[0062] Step 3. Analyze candidate wind turbine WT h The correlation between wind conditions of various wind turbine units is high. Based on the redundancy determination principle, WT is removed. h f1 redundant wind turbines with high correlation to wind conditions and small Euclidean distance.

[0063] Step 4. Select f2 wind turbines with larger feature weights from the remaining wind turbines and add them to the candidate wind turbine WT. h .

[0064] Step 5. Repeat Steps 3 and 4, and analyze the MIC value again until the candidate wind turbine WT is reached. h There are no redundant fans. The final multi-characteristic fan WT = {WT1, WT2, ..., WT} is determined. Q}, where Q represents the number of characteristic fans.

[0065] In summary, the wind condition information of characteristic wind turbines represents a multi-dimensional characteristic wind condition of the wind farm, which not only has a significant impact on wind farm power but is also relatively independent of each other. This method effectively reduces the input dimensionality of the model while ensuring the integrity of the wind farm's wind condition information.

[0066] Step 3: Based on the wind condition selection results of the wind farm, ultra-short-term multi-step prediction of the wind conditions of characteristic wind turbines is achieved through different temporal finite difference (TFD) direct time series modeling strategies. Based on this, a multi-step interval prediction model is established. Specifically, this includes the following: Based on the TFD modeling concept of univariate time series, both multiple inputs single output (MISO)-TFD modeling strategies and multiple inputs multiple outputs (MIMO)-TFD modeling strategies are employed to ensure the reliability and flexibility of wind condition prediction. The univariate time series TFD modeling concept can be understood as the wind conditions from time t+1 to time t+Δt having a certain temporal correlation with the wind conditions before time t+1. A temporal finite difference wind condition vector И(t) is defined. After a finite number of difference operations, the wind conditions Y of the wind farm from time t+1 to time t+Δt can be obtained. wc (t+1, t+Δt), that is:

[0067]

[0068] Where И(t)=[Y wc (t),Y wc (t-1),...,Y wc (tn x )], Y wc (t+1,t+Δt)=[Y wc (t+1),Y wc (t+2),...,Y wc (t+n y )]. Where Y wc For a wind farm, the wind speed or wind direction time series is n. x and n y These represent the input and output orders of the model, respectively. In this case, the Akaike Information Criterion (AIC) is preferred to determine the model order; f(·) represents a finite number of difference operations; e c This is a noise signal.

[0069] Based on the concept of TFD modeling for univariate time series, the multi-input single-output (MISO)-TFD modeling strategy can be described as follows: univariate prediction employs multiple inputs (n... x ≥1) and a single output (n) y=1) Establish a characteristic wind condition prediction model. The model based on the multiple inputs single output (MISO)-TFD modeling strategy can only obtain a single-point prediction value at a time. Through a rolling prediction mechanism, the predicted values ​​are continuously used as model inputs to achieve advanced multi-step prediction. First, the appropriate advanced machine learning neural network algorithm is determined. In this case simulation, a nonlinear auto-regressive (NAR) dynamic neural network and a long short-term memory (LSTM) neural network are used to establish two characteristic wind condition prediction models: NAR_MISO and LSTM_MISO.

[0070] The multi-input single-output (MIMO)-TFD modeling strategy can be described as follows: univariate prediction employs multi-input (n... x ≥1) and multiple outputs (n) y ≥2) Establish a feature-based wind condition prediction model. The model, based on the multiple inputs multiple outputs (MIMO)-TFD modeling strategy, achieves advanced multi-step prediction by establishing a mapping relationship between multi-dimensional inputs and multi-dimensional outputs. Considering the input and output formats, the simulation uses an LSTM neural network to establish an LSTM_MIMO feature-based wind condition prediction model.

[0071] Three characteristic wind condition prediction models—NAR_MISO, LSTM_MISO, and LSTM_MIMO—were established using different time-varying TFD modeling strategies. Wind speed or direction time series data from specific wind turbines were selected to build corresponding wind condition prediction models. The modeling dataset consisted of 3500 sample points, divided into training and testing sets for model training and validation, respectively. Root mean square error (RMSE) and mean percentage error (MSE) were used to evaluate the model prediction results.

[0072] The hyperparameters involved in the simulated LSTM network are: input sequence length I, output sequence length O, number of hidden layers H, and learning rate α. * Learning rate decay coefficient D, number of LSTM layers L, number of iterations N d To optimize the performance of the LSTM network model, the Adam optimization algorithm is employed to ensure fast convergence during training. Furthermore, K-fold cross-validation is used to avoid overfitting. Parameters are continuously adjusted using a trial-and-error method to achieve better modeling results. The NAR network model constructed in the simulation is optimized using the Levenberg-Marquardt backpropagation algorithm. The model's predictive performance is mainly affected by the number of hidden layer neurons, H. d Delay steps D nThe optimal parameters were selected through multiple comparative experiments, taking into account the influence of the input sequence length.

[0073] Based on simulation, appropriate network parameter settings were obtained, and then the wind speed and direction prediction model was trained. A backward prediction of 60 steps was selected to obtain the wind speed and direction prediction results for characteristic wind turbines 1 hour in advance. By comparing the prediction performance of the three models, the model with the better performance was selected as the final characteristic wind condition prediction model to ensure prediction quality.

[0074] After obtaining the results of multi-step predictions, the corresponding interval construction algorithm is determined. In this case, both the nonparametric conditional kernel density estimation (CKDE) method and the semiparametric Copula estimation method are used to establish a multi-step interval prediction model for characteristic wind conditions. Let the output value of the prediction model be Y. * The model error is e. Given Y... * Under these conditions, the conditional probability density function of e can be calculated. The steps for obtaining the multi-step interval model of characteristic wind conditions using the CKDE method are as follows:

[0075] Step 1. Let the predicted characteristic wind condition value be Y. * Let the prediction error be e. Calculate the conditional probability density of e.

[0076] Step 2. Based on Step 1, given a confidence level α, for a given wind speed or wind direction, obtain the fluctuation range of error e based on the upper and lower quantiles of the conditional probability [e]. L ,e H Then compare it with the predicted output value Y. * Combining these, we obtain the uncertainty interval [Y]. L ,Y H ].

[0077] The semi-parametric method uses Copula estimation to obtain the joint distribution function F(Y) of the two variables through the Copula function. * e), the parameter θ is estimated using the maximum likelihood method, and then the output value Y of the known prediction model is obtained. * The probability of error e under given conditions. The steps for obtaining the multi-step interval model of characteristic wind conditions using the Copula method are as follows:

[0078] Step 1. Let the predicted characteristic wind condition value be Y. * The model error is e. Calculate the marginal cumulative distribution F under a given wind speed or wind direction. Y (Y * Then the conditional distribution of error e is obtained.

[0079] Step 2. Based on Step 1, given a confidence level α, calculate the conditional probability quantiles and inversely solve for the confidence interval [e] of the given wind speed or windward error e. L ,e H ] Combined with the predicted output value Y * The uncertainty interval [Y] is obtained L ,Y H ].

[0080] To evaluate the reliability of the interval model, interval coverage probability (ICP) and interval average width (IAW) were used as evaluation indicators. With a confidence level of α = 0.05, interval prediction models for wind speed and direction for each characteristic wind turbine were established based on the CKDE method and the Copula estimation method. Finally, multi-step interval prediction results for wind conditions 1 hour in advance for each characteristic wind turbine were obtained.

[0081] Step 4: Define the finite difference operating domain and implement the convex partitioning of the finite difference computation space. Based on this, implement global temporal finite difference (TFD)-machine learning (ML) modeling of the lumped generation characteristics of the wind farm. Specifically, this includes the following: establishing a lumped output characteristic model of the wind farm based on wind direction and speed prediction results. To effectively characterize the integrated operation mode of the units within the wind farm, a TFD hybrid feature vector φ(t) is defined to classify different wind conditions in the wind farm. φ(t) includes the wind direction and speed of all characteristic wind conditions and covers their time delay relationship with the wind farm's output power. The relationship between the wind farm's power P(t) and φ(t) at time t is as follows:

[0082]

[0083] In the formula, φ(t) = [P(t-1),...,P(tn)] a ),V1(t),V1(t-1),...,V1(tm a ),...,V Q (t),V Q (t-1),...,V Q (tm a ),W1(t),W1(t-1),...,W1(tm a ),...,W Q (t),W Q (t-1),...,W Q (tm a ]], where P is the output power of the wind farm, and V and W are the characteristic wind turbine wind speed and wind direction; na and m a The model order is determined using the AIC criterion.

[0084] The finite difference operating domain χ is formed by the space spanned by the TFD mixed feature vector φ(t), and high-dimensional clustering and hyperplane partitioning are used to identify the operating domain.

[0085] The power and wind speed in the TFD hybrid eigenvectors are normalized to the interval [-1, 1], forming the variable Φ1(t) = {x i |i=1,2,...,N}, where N is the number of samples, x i ={x i1 ,x i2 ,...,x ih}, h is the sample x i Dimension. Let the wind direction component variable in the TFD mixed eigenvector be Φ2(t)={w i For any two wind direction samples |i=1,2,...,N}, the formula for calculating the angle of change of wind direction is:

[0086]

[0087] In the formula w i ={w i1 ,w i2 ,...,w iM},w j ={w j1 ,w j2 ,...,w jM}, M is the dimension, d = 1, 2, ..., M, and W wd It lies within the interval [-1, 1].

[0088] Based on the above, the normalized TFD mixture feature vector φ is obtained. norm (t)=[Φ1(t),Φ2(t)], for any two samples, calculate the new similarity metric:

[0089]

[0090] Where φ norm,i (t)=[x i1 ,x i2 ,...,x ih ,w i1 ,w i2 ,...,w iM ],φ norm,j (t)=[x j1 ,x j2 ,...,x jh ,w j1,w j2 ,...,w jM ], i, j = 1, 2, ..., N.

[0091] After clustering, the subset D = {D} is obtained. s |s=1,2,...,C}, where C is the number of clusters. To further evaluate the clustering effect, and to improve the traditional clustering evaluation index SSE, a new clustering effect evaluation index is defined:

[0092]

[0093] Where φ norm,s (t) is the cluster center, φ norm,R (t) represents the subset D s The remaining samples in. φ norm,s (t)=[x s1 ,x s2 ,...,x sh ,w s1 ,w s2 ,...,w sM ],φ norm,R (t)=[x R1 ,x R2 ,...,x Rh ,w R1 ,w R2 ,...,w RM ].

[0094] Clustering divides φ(t) into C subsets, with the feature vectors in each subset forming a sub-operational domain. Solving for the sub-operational domain can be transformed into finding the switching hyperplane between the subsets. The equation of the switching hyperplane is: H ij ={wφ T (t)+F=0}, where w is the normal vector and F is the offset. A soft-margin support vector machine is used to construct the objective function and solve for the hyperplane coefficients, enabling rapid automatic switching of data between sub-operational domains.

[0095] Based on the defined operating domains, sample balance is assessed, and samples are selected for unified modeling and domain-specific modeling. First, samples are selected for each operating domain. Then, machine learning algorithms and cross-validation mechanisms are chosen to establish a unified model and domain-specific models. The former constructs the input-output mapping relationship across the entire finite difference domain, while the latter establishes input-output mapping models separately for each sub-operating domain. If the accuracy of the two is similar, it indicates that the machine learning algorithm has sufficient global learning capability, and the unified model is sufficient. If the latter has higher accuracy, it indicates that the nonlinearity of the modeling process is too complex, exceeding the learning capability of the machine learning algorithm, and it is recommended to use the domain-specific model. In summary, the model with the best overall performance is selected as the final modeling scheme for the lumped-out power output characteristics of the wind farm.

[0096] Construct a time-finite-difference regression vector based on the wind speed of all characteristic wind conditions and the time delay relationship between it and wind power. Finite difference regression vector The wind farm output power P(t) is used as the model input and as the model output. An LSTM neural network is used to establish the power output characteristic model. During the simulation, the Adam optimization algorithm and ten-fold cross-validation are used. The parameter values ​​are continuously changed through trial and error to obtain good modeling results. The normalized root mean square error (NRMSE) and mean absolute percentage error (MAPE) are used to measure the model accuracy.

[0097] Based on the above parameter setting method, the unified model and the domain-specific model were trained and optimized. Test sets under different working conditions were selected to verify the model effect. The final modeling scheme was determined according to the complexity of the selected wind farm. Since the sampling interval of the wind farm in the selected case is 1 minute, the wind power generation characteristics are more volatile. Therefore, based on the simulation results, the domain-specific model with higher model accuracy was selected as the final ultra-short-term wind power generation characteristic modeling scheme to cope with the complex nonlinearity and uncertainty of the wind farm.

[0098] Step 5: Based on the established characteristic wind condition prediction model and wind farm lumped generation characteristic model, and using a rolling prediction mechanism, realize ultra-short-term multi-step interval prediction of the wind farm. Specifically, this includes the following: First, obtain the ultra-short-term multi-step power prediction values ​​of the wind farm. The specific steps are as follows: First, based on the characteristic wind condition prediction model, perform multi-step predictions of wind speed and direction within the target prediction period. Based on the wind speed and direction prediction information, generate a TFD hybrid feature vector φ(t). Determine the operating domain to which φ(t) belongs based on the calculated hyperplane equation. Then, select the corresponding wind power output characteristic model, which can be a unified model or a domain-specific model. Second, input the multi-step interval prediction values ​​of wind speed into the selected global TFD-ML model, and obtain the ultra-short-term multi-step prediction results of wind power based on the rolling prediction mechanism. Finally, conduct online evaluation and monitoring of the prediction accuracy. If the prediction accuracy continues to be lower than the set requirements, the corresponding TFD-ML model needs to be updated.

[0099] The second step is to obtain a multi-step interval model of the wind farm's ultra-short-term power. Based on the obtained multi-step wind power prediction values, a multi-step interval prediction model of wind power is established using nonparametric CKDE and semiparametric Copula estimation. The method for obtaining the multi-step interval prediction model based on the CKDE and Copula methods has been introduced in step three; simply let the multi-step wind power prediction value be Y. * With a prediction error of e, the predicted wind power value Y can be obtained. * Uncertainty interval [Y] L ,Y H ].

[0100] During the simulation, the wind speed and direction prediction results for the target time period 1 hour in advance are first obtained, φ(t) is constructed, and its operating domain is determined. A prediction step size of 60 is chosen, and by continuously inputting the corresponding wind speed and power prediction values ​​into the wind farm power generation characteristic model, the wind farm power prediction value for 1 hour in advance is obtained. With a confidence level of α = 0.05, a multi-step interval prediction model for wind power 1 hour in advance for the target time period is established using the obtained multi-step wind power prediction value Y* and the multi-step prediction error e. The final prediction results of the selected case in this invention are shown in the appendix. Figure 2 The evaluation indicators for the prediction effect are shown in Table 1.

[0101] Table 1 Calculation results of evaluation indicators for the prediction model

[0102]

[0103] Since output wind power is highly dependent on input wind energy, this invention adopts a stepwise prediction approach, proposing a method for ultra-short-term interval prediction of wind power based on characteristic wind condition selection and temporal finite difference dynamic modeling. First, considering the spatiotemporal dispersion of wind conditions, a multi-dimensional characteristic wind condition of the wind farm is selected based on wind farm operation data and geographic information to reduce model complexity. Then, based on the time dependence of wind resource prediction, a multi-step ultra-short-term interval prediction model for characteristic wind turbine wind conditions is established using different temporal finite difference (TFD) direct time series modeling strategies. The accuracy of characteristic wind condition prediction is ensured by comparing the effects of different modeling strategies. Second, given the complex wind conditions and time delays between outputs in wind farms, a finite difference operating domain is constructed. The proposed global temporal finite difference (TFD)-machine learning (ML) method is used to model the comprehensive power generation characteristics of the wind farm between the dispersed multiple wind conditions and the total power of the wind farm. Based on this, a flexible ultra-short-term wind power prediction scheme for wind farms can be established. Ultimately, a multi-step interval prediction model for wind power is established using conditional kernel density estimation (CKDE) and Copula estimation. Combining modular thinking to achieve systematic wind power prediction has significant guiding implications for the optimal scheduling of new power systems.

[0104] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely principles of the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the claimed invention. The scope of protection claimed by the appended claims and their equivalents is defined.

Claims

1. A modular method for multi-step interval prediction of ultra-short-term wind farm power, characterized in that, Includes the following steps: Step 1: Obtain the geographical location information of the wind farm, and complete the collection and storage of historical operation data of the wind farm at a certain sampling interval; Step 2: Based on the geographical information and historical operation data of the wind farm, select wind turbines with characteristics of the wind farm and obtain the multi-characteristic wind conditions of the wind farm; Step 3: Based on the wind farm characteristic wind condition selection results, establish a multi-step interval prediction model for characteristic wind conditions through different time finite difference (TFD) direct time series modeling strategies; This study employs a univariate time series-based finite-difference (TFD) modeling strategy, while simultaneously utilizing a multi-input single-output (MIMO-TFD) modeling strategy to ensure the reliability and flexibility of wind condition forecasting. The MIMO-TFD modeling strategy uses machine learning neural network algorithms to establish a feature-based wind condition prediction model. Through a rolling prediction mechanism, the predicted values ​​are used as model inputs to achieve multi-step advance prediction of feature-based wind conditions. Furthermore, the MIMO-TFD modeling strategy uses machine learning neural network algorithms to establish a mapping relationship between multi-dimensional inputs and multi-dimensional outputs, enabling multi-step advance prediction of feature-based wind conditions. The multi-step interval prediction model for feature-based wind conditions, based on the multi-step prediction results, utilizes both non-parametric and semi-parametric interval construction algorithms to establish a multi-step interval prediction model, effectively ensuring the reliability of the uncertain interval model. Step 4: Define the finite difference operating domain and implement the convex partitioning of the finite difference operation space. Based on this, implement global time finite difference-machine learning (TFD-ML) modeling of the lumped generation characteristics of wind farms. Step 5: Based on the established characteristic wind condition prediction model and wind farm lumped generation characteristic model, and based on the rolling prediction mechanism, realize ultra-short-term multi-step interval prediction of wind farm; Step 4, which involves global time finite difference-machine learning (TFD-ML) modeling of the lumped generation characteristics of wind farms, specifically includes the following steps: Step 1. Based on the partitioning results of finite difference operations, select samples in each sub-operational domain; Step 2. Select time series modeling machine learning algorithms and cross-validation mechanisms to establish a unified model and sub-operational domain models; the unified model is to construct the mapping relationship between input and output in the entire finite difference domain, and the sub-operational domain models are to establish the mapping model between input and output in each sub-operational domain. Step 3. Analyze the complexity of the selected wind farm operation: If the accuracy of the two models is similar, it indicates that the machine learning algorithm has sufficient global learning ability, and a unified model can be used; if the accuracy of the sub-operation domain model is higher than that of the unified model, it indicates that the nonlinearity of the modeling process is too complex and exceeds the learning ability of the machine learning algorithm, and the sub-operation domain model should be used. Step 4. Select the model with the best performance as the final modeling scheme for the lumped generation characteristics of the wind farm, further improving the flexibility of modeling and effectively addressing the complex nonlinearity and uncertainty of the wind farm. The unified model and the sub-operational domain model use a finite difference regression vector (t) as the model input and the wind farm output power P(t) as the model output for modeling. The finite difference regression vector (t) considers the most direct energy conversion relationship between wind speed and wind power. Based on the wind speed of all characteristic wind conditions and its time delay relationship with wind power, a time finite difference regression vector is defined: Based on the wind speed of all characteristic wind conditions and the time delay relationship between it and wind power, a time finite difference regression vector φ(t) = [P(t-1), ...,P(tn)] is defined. a ),V1 (t), V1 (t-1),...,V1 (tm a ),...,W Q (t),W Q (t-1), ...,W Q (tm a )]; Where P is the output power of the wind farm, and V and W are the characteristic wind turbine wind speed and wind direction; n a and m a The model order is determined by selecting the appropriate model order determination algorithm.

2. The modular method for multi-step interval prediction of ultra-short-term wind farm power according to claim 1, characterized in that: Step 1 further includes the following: the geographical location information of the wind farm includes the longitude, latitude, and altitude information of all wind turbines within the wind farm; the sampling interval must meet the requirements of ultra-short-term wind power prediction, and the sampling time interval defined for ultra-short-term power prediction is 1 to 15 minutes; the historical operating data of the wind farm includes historical data on the active power, wind speed, and wind direction of all wind turbines in the wind farm, and the selected historical operating data of the wind farm must be processed for abnormal data cleaning in order to provide a data sample basis for subsequent ultra-short-term wind power prediction.

3. The modular method for multi-step interval prediction of ultra-short-term wind farm power according to claim 1, characterized in that: Step 2 further includes the following: The selection method for the multi-characteristic wind conditions of the wind farm includes two parts: an external correlation analysis method and an internal correlation analysis method. First, an external correlation analysis is performed using a feature selection algorithm. By establishing a feature selection model, candidate wind turbines WTh for the wind farm are determined. Secondly, combining the geographical location information of the wind farm and using a nonlinear correlation analysis algorithm for internal correlation analysis, redundant wind turbines in the candidate wind turbines are eliminated. Finally, the wind conditions of the characteristic wind turbines obtained through the wind farm characteristic wind turbine selection method are the multi-dimensional characteristic wind conditions of the wind farm. The feature selection algorithm is as follows: it simultaneously considers the influence of wind speed and wind direction on the power of the wind farm. The external correlation analysis method specifically includes the following: a) If the correlation between wind speed and wind direction is strong, a dimensionality reduction algorithm is used to reduce the dimensionality of the original wind direction and wind speed data, and the dimensionality-reduced data Ddim and normalized wind farm power are used as the input and output of the established feature selection model, respectively. According to the feature weights returned by the feature selection model, the wind turbines with larger feature weights are selected as candidate wind turbines WTh. b) If the correlation between wind speed and wind direction is weak, and the original information component contained in the data after dimensionality reduction using the corresponding dimensionality reduction algorithm is small, it indicates that wind speed and wind direction are two relatively independent variables. Two feature selection models are established separately to calculate the impact of wind speed and wind direction on wind farm power separately. Wind turbines with larger feature weights in both models are selected as candidate wind turbines WTh. The internal correlation analysis specifically involves: quantifying the geographical location information within the wind farm and calculating the nonlinear correlation of wind conditions between wind turbine units to eliminate redundant turbines; the method for quantifying the geographical location information within the wind farm specifically includes the following: normalizing the longitude, latitude, and altitude information of each operating unit within the wind farm and then calculating the Euclidean distance between each pair of units; the method for nonlinear correlation analysis of wind conditions between wind turbine units specifically includes the following: determining the nonlinear correlation analysis algorithm, and performing wind condition correlation analysis in two cases based on the different strengths of the correlation between wind speed and wind direction: (1) If the wind speed and wind direction of the wind turbine are strongly correlated, calculate the nonlinear correlation between the dimensionality-reduced data Ddim of each wind turbine. (2) If the correlation between wind speed and wind direction of wind turbine is weak, calculate the nonlinear correlation between the original wind speed and the original wind direction of each wind turbine. The method for eliminating redundant wind turbines specifically includes the following: make a comprehensive judgment based on the normalized Euclidean distance of wind turbines and the wind condition correlation analysis results. If the Euclidean distance between two wind turbines is small and the wind condition correlation is high, then they are considered to be geographically close and have similar wind condition information. They are determined to be redundant wind turbines. Based on the feature selection model results in the external correlation analysis, wind turbines with smaller feature weights are eliminated, thus eliminating redundant wind turbines. Combining the aforementioned external correlation analysis and internal correlation analysis, the wind farm characteristic turbine selection method specifically includes the following steps: Step 1. Select candidate wind turbines WTh through external correlation analysis; Step 2. Quantify the geographical location information within the wind farm and complete the calculation of the nonlinear correlation of wind conditions for each wind turbine. Step 3. Analyze the wind condition correlation of each wind turbine in WTh. Based on the redundancy determination principle, remove f1 redundant wind turbines with high wind condition correlation and small Euclidean distance. Step 4. Select f2 wind turbines with larger feature weights from the remaining wind turbine units to add as candidate wind turbines WTh; Step 5. Repeat Step 3 and Step 4 to analyze the strength of the nonlinear correlation of wind conditions in wind turbine units again until there are no redundant wind turbines in WTh; finally determine the multivariate characteristic wind turbine WT={WT1,WT2,...,WTQ}, where Q represents the number of characteristic wind turbines.

4. The modular method for multi-step interval prediction of ultra-short-term wind farm power according to claim 1, characterized in that: Step 4 further includes the following: convex partitioning of the finite difference operation space, specifically including two parts: high-dimensional clustering and hyperplane calculation. The high-dimensional clustering specifically includes the following: considering the periodic wind direction data in φ(t), based on the partitioned clustering algorithm process, redefining the similarity metric d(φnorm,i(t),φnorm,j(t)) and the clustering effect evaluation metric OSSE to complete the classification of the data; The calculation method for the similarity metric d(φnorm,i(t),φnorm,j(t)) includes the following steps: Step 1. Let the wind direction component variable Φ2(t) = {wi|i =1,2, ..., N} in the TFD mixed feature vector φ(t), where N is the number of samples. For any two wind direction samples wi and wj, calculate the wind direction change angle Wwd to complete the preprocessing of the periodic wind direction data. Step 2. Normalize the power and wind speed in the TFD mixed feature vector φ(t) to the interval [-1, 1] to form the variable Φ1(t) = {xi | i = 1, 2, ..., N}, where N is the number of samples; define the normalized TFD mixed feature vector φnorm(t) = [Φ1(t), Φ2(t)]. For any two samples, the new similarity metric is calculated as follows: Where φnorm, i(t) = [xi1, xi2, ..., xih, wi1, wi2, ..., wiM], φnorm, j(t) = [xj1, xj2, ..., xjh, wj1, wj2, ..., wjM]; M is the dimension of the wind direction data w; h is the dimension of the variable x, i, j = 1, 2, ..., N; The specific calculation formula for the clustering performance evaluation index OSSE is as follows: Where Ds is the subset of data obtained after clustering; C is the number of clusters; φnorm,s(t) is the cluster center; and φnorm,R(t) is the remaining sample points in the subset Ds after removing the cluster centers.

5. The modular method for multi-step interval prediction of ultra-short-term wind farm power according to claim 1, characterized in that: step 5 further includes the following: the implementation steps for ultra-short-term multi-step interval prediction of wind farms are as follows: Step 1. Based on wind speed and wind direction prediction information, generate a TFD hybrid feature vector φ(t); Step 2. Based on the convex partitioning result of the finite difference operation space, determine its operating domain and select the corresponding prediction model, adopting a unified model or a sub-operating domain model; Step 3. Input the ultra-short-term interval prediction results of wind speed into the global TFD-ML model to obtain the ultra-short-term multi-step prediction results of wind farm output power; Step 4. Based on the multi-step prediction results of wind power in wind farms, determine the prediction interval construction algorithm. At the same time, non-parametric and semi-parametric interval construction algorithms are used to establish an ultra-short-term wind power multi-step interval prediction model to effectively ensure the reliability of the uncertain interval model. Step 5. Evaluate and monitor the prediction accuracy online. If the prediction accuracy continues to be lower than the set requirements, the TFD-ML model needs to be updated.

6. A non-volatile storage medium, characterized in that, The non-volatile storage medium includes a stored program, wherein the program, when executed, controls the device where the non-volatile storage medium is located to perform the method described in any one of claims 1 to 5.

7. An electronic device, characterized in that, It includes a processor and a memory; the memory stores computer-readable instructions, and the processor is configured to execute the computer-readable instructions, wherein the computer-readable instructions, when executed, perform the method according to any one of claims 1 to 5.

Images