Adaptive prediction method for wind turbine power generation under non-stationary data flow

Abstract

The present invention discloses a method for adaptively predicting wind turbine power generation under non-stationary data streams, comprising the following steps: S1. extracting features from numerical weather forecast (NWP) data to obtain sample-related features and sequence-related features, forming a sample-sequence feature vector; S2. performing similarity search and clustering based on the sample-sequence feature vector to divide different meteorological modes; S3. offline training a LightGBM network using historical NWP and power data for each meteorological mode to obtain a power prediction model, performing power prediction based on real-time collected NWP meteorological data, monitoring the prediction performance of the power prediction model under each meteorological mode in real time based on drift detection and DDA algorithms, and adaptively adjusting model parameters to achieve adaptive prediction of wind power under non-stationary data streams. The present invention can monitor changes in the input-output mapping relationship in real time and adaptively adjust the model to adapt to the new mapping relationship, thereby improving power prediction accuracy.

Description

Technical Field

[0001] The present invention belongs to the technical field of wind power prediction, and in particular relates to a method for adaptively predicting the generated power of a wind turbine under non-stationary data flow. Background Art

[0002] Predicting wind turbine power generation is crucial for increasing wind farm profitability. Due to the large, random fluctuations in wind power, its intermittent and uncertain nature frequently disrupts the power grid. Therefore, accurately predicting future power generation, enabling timely adjustment of controllable turbines, and ensuring grid stability are prerequisites for large-scale wind power integration. Furthermore, wind power exhibits the characteristic of peak load regulation, with power generation often concentrated at night and less during the day. This is contrary to the peak-valley distribution of electricity load.

[0003] To ensure an instantaneous balance between power generation and load, the power grid develops dispatch plans in advance for wind farm technicians to implement. Wind power forecasts, a key input variable in scheduling, indirectly determine how much wind turbine power can be effectively absorbed and are a key factor influencing wind farm revenue.

[0004] Currently, wind power forecasting methods use historical wind speed, wind direction, or other meteorological data as input to train model parameters and predict future power. The model's structure and parameters are typically fixed after training, assuming that the mapping relationship between generated power and input variables learned from historical data will remain valid for future power generation scenarios. However, due to the non-stationarity of meteorological data such as wind speed, the distribution of input variables changes over time, and the mapping relationship between input and output also varies. This dynamic non-stationarity in data distribution causes the accuracy of traditional offline models to gradually decrease over time.

[0005] The prediction methods in the existing technology are unable to adapt to the non-stationarity of the input data, resulting in a decrease in prediction accuracy over time. They are unable to accurately predict the power generation of wind turbines, and are unable to maintain the accuracy and adaptability of the prediction model under constantly changing meteorological conditions and wind turbine status. It is difficult to meet the needs of wind farm and power grid operators for high-precision power prediction. Summary of the Invention

[0006] The technical problem to be solved by the present invention is to address the deficiencies of the above-mentioned existing technologies and provide an adaptive prediction method for wind turbine power generation under non-stationary data streams, which can monitor changes in the input-output mapping relationship in real time and adaptively adjust the model to adapt to the new mapping relationship, thereby improving the power prediction accuracy.

[0007] In order to achieve the above technical objectives, the technical solution adopted by the present invention is:

[0008] The method for adaptively predicting the generated power of a wind turbine under non-stationary data flow includes the following steps:

[0009] S1. Extract features from numerical weather forecast (NWP) data to obtain sample-related features and sequence-related features, and construct a sample-sequence feature vector.

[0010] S2. Perform similarity search and clustering based on sample-sequence feature vectors to identify different meteorological patterns;

[0011] S3. Use the historical NWP and power data of each meteorological mode to train the LightGBM network offline to obtain the power prediction model. Perform power prediction based on the real-time collected NWP meteorological data. Based on the drift detection and DDA algorithm, monitor the prediction performance of the power prediction model under each meteorological mode in real time, and adaptively adjust the model parameters to achieve adaptive prediction of wind power under non-stationary data streams.

[0012] To optimize the above technical solutions, specific measures taken also include:

[0013] The above-mentioned S1 includes the following sub-steps:

[0014] S11. Use a sliding window to divide the original input NWP data of each meteorological variable into data segments of equal length. For meteorological variable i, the data segment of length T in the t-th time window is The meteorological variables include wind speed, wind direction sine, wind direction cosine, temperature, air pressure, and air density;

[0015] S12. Extract 14 sample-related features from the data segment, including the arithmetic mean, maximum value, minimum value, 5th and 95th percentiles, first, second, and third quartiles, number of occurrences of minimum and maximum values, range, variance, approximate unbiased estimate of skewness, and approximate unbiased estimate of kurtosis of meteorological variable i in time window t;

[0016] S13. Extracting 19 sequence-related features from the data fragment, including:

[0017] Autocorrelation related features: autocorrelation function ACF, partial autocorrelation function PACF, Box-Pierce statistic, Hurst index;

[0018] Volatility-related characteristics: Detrended volatility analysis (DFA), Lomb-Scargle periodogram, level shift, variance change, crossover point, Shannon entropy, Kullback-Leibler divergence;

[0019] Time series decomposition related features: trend factor, seasonal factor, detrended series correlation, Spikiness intensity, Lumpiness intensity;

[0020] Time-related features: the time of occurrence of maximum value, minimum value, and maximum KL divergence;

[0021] S14. Normalize the extracted sample-related features and sequence-related features, and combine the features of the six meteorological variables to form a 198-dimensional sample-sequence feature vector of the data segments divided by the same time window.

[0022] The trend factors mentioned above are:

[0023]

[0024] The seasonal factors are:

[0025]

[0026] Where Var(·) is the variance function; is the deseasonal sequence; is a detrended series; is a stable normalized time series after Box-Cox transformation; respectively The trend component, seasonal component and residual component obtained by LOESS decomposition;

[0027] The detrended correlation is:

[0028]

[0029] Where, is the autocorrelation function of the detrended NWP series i at the kth lag in the time window t, n is the length of the series, and h is the maximum lag considered.

[0030] The above-mentioned S2 includes the following sub-steps:

[0031] S21. Perform similarity search based on the sample-sequence feature vector by calculating the Euclidean distance;

[0032] S22. The k-means++ algorithm is used to cluster data segments based on similarity search results. The optimal number of meteorological patterns is determined by the elbow rule, and similar data segments are clustered into different meteorological patterns.

[0033] The above S3 includes the following sub-steps:

[0034] S31. For each meteorological pattern, use historical NWP data and its corresponding historical power data to train the LightGBM network offline to obtain the power prediction model for the corresponding meteorological pattern;

[0035] S32. Collect NWP data in real time, extract sample-sequence feature vectors, calculate the similarity between the sample-sequence feature vectors and each meteorological model, match the corresponding meteorological model, and input the real-time collected NWP data into the power prediction model of the corresponding meteorological model for power prediction;

[0036] S33. Use drift detection and DDA algorithm to divide the power prediction model into three states: normal, warning, and drift through the warning threshold Tp1 and the drift threshold Tp2. Monitor the prediction performance of the model in real time, and decide whether to update the model parameters according to the change of prediction accuracy in different states, so as to realize the adaptive adjustment of model parameters.

[0037] The above-mentioned S33 uses drift detection and DDA algorithm to divide the power prediction model into three states: normal, warning, and drift through the warning threshold Tp1 and the drift threshold Tp2, which is specifically as follows:

[0038] The input NWP data is input into the model through a sliding window, and the prediction accuracy decrease value of the model for two adjacent windows is calculated:

[0039] Dif p = RMSE1 / RMSE2<00001�5>

[0040] where RMSE1 and RMSE2 are the sum of the root mean squares of the model prediction errors for two adjacent windows;

[0041] When Dif p < Tp1, the model remains in the normal state. When Tp2 ≥ Dif p > Tp1, the model enters the warning state. When Dif p > Tp2, the model enters the drift state.

[0042] The above-mentioned S33 decides whether to update the model parameters according to the change of prediction accuracy in different states, so as to realize the adaptive adjustment of model parameters, including:

[0043] 1) In the normal state, if the prediction accuracy decrease value of the current LightGBM model for two adjacent windows satisfies Tp2 ≥ Dif p > Tp1, it enters the warning state, and the cache window starts to collect new samples;

[0044] 2) In the warning state, if the prediction accuracy further decreases, and Dif p > Tp2 and the duration of the warning state does not exceed the set duration, the model is retrained with new samples and enters the drift state, and the performance benchmark is recorded at the same time; if the prediction accuracy further decreases, but Tp2 ≥ Dif p > Tp1 and the duration of the warning state exceeds the set duration, it is regarded as a false alarm and returns to the normal state;

[0045] 3) In the drift state, the cache window continues to collect samples and continuously evaluates the performance of the current model on the newly collected samples. When RMSE1 / Bas p >Tp1, Bas p To establish a new concept’s performance benchmark, the model is retrained using samples in the cache window and the cache window is released after the status returns to normal.

[0046] The present invention has the following beneficial effects:

[0047] The present invention extracts sample-related and sequence-related features from numerical weather forecast (NWP) data segments to form a 198-dimensional feature vector. Based on the k-means++ algorithm, the feature vectors are clustered according to their similarity to divide meteorological patterns. The DDA-LightGBM adaptive prediction framework is used to train the model, and the DDA algorithm is used to monitor the prediction performance in real time. The model parameters are dynamically updated through the threshold division state to cope with the drift of the input-output mapping relationship. This method can effectively improve the prediction accuracy and robustness and optimize the power generation revenue of the wind farm.

[0048] The present invention solves the problem of decreased accuracy of traditional offline models caused by non-stationary meteorological data in wind power forecasting by real-time monitoring and adjustment of model parameters. Through the proposed sample-sequence feature extraction, meteorological pattern matching and drift detection and adaptation-based strategies, the wind power forecasting accuracy is significantly improved, the non-stationary changes of data are effectively dealt with, the model adaptability and computational efficiency are enhanced, and the excellent practical value is demonstrated in the optimized operation of wind farms and stable grid dispatching. BRIEF DESCRIPTION OF THE DRAWINGS

[0049] Figure 1 is an adaptive power prediction flow chart of the present invention;

[0050] Figure 2 is a sample-sequence feature graph of the present invention;

[0051] Figure 3 is a similarity search result graph under different feature spaces of the present invention;

[0052] Figure 4 1 is a graph showing the prediction results of different models of the present invention in segment 1;

[0053] Figure 5 2 is a graph showing the prediction results of different models of the present invention in segment 2;

[0054] Figure 6 3 is a graph showing the prediction results of different models of the present invention in segment 3;

[0055] Figure 7 4 is a graph showing the prediction results of different models of the present invention in segment 4;

[0056] Figure 8 This is the RMSE graph of multi-step predictions of different models in the test set of the present invention. DETAILED DESCRIPTION

[0057] In order to make the purpose, technical solutions and advantages of the present invention more clearly understood, the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention.

[0058] Although the steps in the present invention are arranged with numbers, they are not intended to limit the order of the steps. Unless the order of the steps is clearly stated or the execution of a step requires other steps as a basis, the relative order of the steps can be adjusted. It is understood that the term "and / or" used herein refers to and covers any and all possible combinations of one or more of the associated listed items.

[0059] The present invention provides a method for adaptively predicting the generated power of a wind turbine under non-stationary data flow, comprising the following steps:

[0060] S1. Extract features from numerical weather forecast (NWP) data to obtain sample-related features and sequence-related features, and construct a sample-sequence feature vector.

[0061] S2. Perform similarity search and clustering based on sample-sequence feature vectors to identify different meteorological patterns;

[0062] S3. Use the historical NWP and power data of each meteorological mode to train the LightGBM network offline to obtain the power prediction model. Perform power prediction based on the real-time collected NWP meteorological data. Based on the drift detection and DDA algorithm, monitor the prediction performance of the power prediction model under each meteorological mode in real time, and adaptively adjust the model parameters to achieve adaptive prediction of wind power under non-stationary data streams.

[0063] In the embodiment, S1 performs sample-sequence feature extraction, extracts features from the input NWP data, and obtains sample-related features and sequence-related features. It first segments the historical NWP data and extracts sample-sequence feature vectors therefrom, including the following sub-steps:

[0064] S11. Use a sliding window to divide the original input data into segments of equal length. For variable i, the tth data segment has a length of T (T = 24, corresponding to 2 hours)

[0065] S12. Extract 14 sample-related features from a given data segment, including the following: the arithmetic mean, maximum value, minimum value, 5th and 95th percentiles, first, second, and third quartiles, the number of occurrences of the minimum and maximum values, and approximately unbiased estimates of the range, variance, skewness, and kurtosis of meteorological variable i in the tth time window;

[0066] S13. Extract 19 serial correlation features, including autocorrelation-related features (autocorrelation function (ACF), partial autocorrelation function (PACF), Box-Pierce statistic, Hurst exponent), volatility-related features (detrended volatility analysis (DFA), Lomb-Scargle periodogram, level shift, variance change, crossover point, Shannon entropy, Kullback-Leibler divergence), time series decomposition-related features (trend factor, seasonal factor, detrended serial correlation, Spikiness intensity, Lumpiness intensity), and occurrence time-related features (maximum value, minimum value, and time of occurrence of maximum KL divergence);

[0067] Stable normalized time series after Box-Cox transformation Perform LOESS decomposition:

[0068]

[0069] in, Corresponding to the trend component, seasonal component and residual component respectively. Therefore, the detrended series is calculated as The deseasonal series is calculated as

[0070] Trend factor: Use It represents the overall level of growth or decrease of the time series and is defined as:

[0071]

[0072] Seasonal factor: Use It represents the periodic variation of the time series and is defined as:

[0073]

[0074] Detrending correlation: Use denoted by , is the serial correlation measured using the Box-Pierce test and is defined as:

[0075]

[0076] Where, is the autocorrelation function of the detrended NWP series i at the kth lag in the time window t, n is the length of the series, and h = 6 is the maximum lag considered.

[0077] S14. Normalize all extracted features and combine the features of the six NWP variables (wind speed, wind direction sine, wind direction cosine, temperature, air pressure, and air density) to form a 198-dimensional sample-sequence feature vector.

[0078] In S1, the sample-sequence feature extraction includes features such as Figure 2 shown.

[0079] In the embodiment, S2 performs meteorological pattern matching, clusters the extracted feature vectors, and divides different meteorological patterns;

[0080] Calculate the similarity between each vector and perform clustering based on k-means++ to divide the typical meteorological patterns and achieve meteorological pattern matching, including the following sub-steps:

[0081] S21. Perform similarity search based on the sample-sequence feature vector by calculating the Euclidean distance;

[0082] S22. Cluster the data segments based on the similarity search results using the k-means++ algorithm, determine the optimal number of meteorological modes using the elbow rule, and cluster similar NWP data segments into different meteorological modes.

[0083] In the embodiment, S3 performs power prediction based on drift detection and adaptation, monitors the prediction performance under each meteorological mode in real time, and adaptively adjusts model parameters.

[0084] Power prediction based on drift detection and adaptation includes the following sub-steps:

[0085] S31. Offline prediction model training based on LightGBM: For each meteorological mode identified in the previous step, a LightGBM model is trained offline using historical NWP and power data as the initial prediction model. The LightGBM model takes a 198-dimensional sample-sequence feature vector as input and outputs power five minutes later.

[0086] S32. Weather pattern matching: Real-time NWP data is collected, sample-sequence feature vectors are extracted, the similarity between the sample-sequence feature vectors and each weather pattern is calculated, the corresponding weather pattern is matched, and the real-time NWP data is input into the power prediction model of the corresponding weather pattern for power prediction;

[0087] The similarity is achieved based on the Euclidean distance of the NWP sequences:

[0088]

[0089] In the formula, V is the feature space, and v q,i and v s,i represent the time series q and s respectively. A smaller d s,q value means that the sequence s has a pattern highly similar to the sequence q. In this method, the NWP sequence has been transformed into V t , so the subsequent similarity distance calculation is based on the feature vector V t after feature extraction.

[0090] S33. Adaptive power prediction. In the current mode, after predicting the future power using the existing prediction model, the drift detection and adaptation (DDA) algorithm is used to divide the power prediction model into three states: normal, warning, and drift through two thresholds (Tp1, Tp2), monitor the prediction performance of the model in real time, and decide whether to update the model parameters according to the change in prediction accuracy in different states to achieve adaptive adjustment of the model.

[0091] For each meteorological mode, the initial LightGBM model is trained using historical power and NWP data. The input NWP data is input into the model through the sliding window W to detect drift and update the LightGBM model. The whole algorithm process has three states: normal state, warning state, and drift state.

[0092] In the normal state, a decrease in the prediction accuracy of the current LightGBM model M for two adjacent windows indicates that there may be a drift in the input-output mapping relationship. Here, the decrease in prediction accuracy is recorded as: [[ID=2​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​

[0096] Specifically, based on the algorithm in Table 1, the drift of the input-output mapping relationship is monitored in real time, and the model is automatically updated to adapt to the new mapping mode.

[0097] Table 1 Overall process of DDA algorithm

[0098]

[0099]

[0100] The specific steps to achieve model adaptive adjustment are:

[0101] 1) Under normal conditions, if the prediction accuracy of the current LightGBM model for two adjacent windows decreases (Tp2≥Dif p >Tp1), enters the warning state, and the cache window begins to collect new samples;

[0102] 2) In the early warning state, if the prediction accuracy is further reduced (Dif p >Tp2), the model is retrained with new samples and enters the drift state, while recording the performance benchmark; if the prediction accuracy does not exceed the set threshold (Tp2≥Dif p >Tp1) and the warning state lasts too long, it is considered a false alarm, the warning is cleared and the normal state is restored;

[0103] 3) In the drift state, the cache window continues to collect samples. If the new model returns to the warning state, that is, when RMSE(M(W i ),y i ) / Bas p >Tp1, Bas p To establish a new concept’s performance benchmark, the model is retrained using samples in the cache window and the cache window is released after the status returns to normal.

[0104] During the drift state, the performance of the current model on newly collected samples is continuously evaluated. By calculating the error metrics (such as the root mean square error (RMSE)) between the model's predictions and the actual observations, the changing trends of the model's performance are analyzed. This helps determine how well the model adapts to the new data distribution and whether an immediate model update is needed.

[0105] During the forecasting process, new NWP data segments first pass through the meteorological pattern matching module to calculate their similarity with each meteorological pattern and assign them to the corresponding meteorological pattern.

[0106] Power forecasting is performed using an adaptive forecasting model (DDA-LightGBM) that matches this meteorological pattern, while forecast performance is continuously monitored using the DDA algorithm. Based on changes in forecast accuracy, model parameters are adjusted in real time according to the three preset state transition rules to ensure the model can adapt to non-stationary changes in input data and maintain excellent forecast accuracy.

[0107] In actual applications, the system uses two hours of historical NWP data as input to predict wind power for the next five minutes. The system updates the predicted power every five minutes and applies it to wind farm power forecasting to assist the power grid in formulating dispatching plans, ensuring a balance between the power generation and load ends, and increasing wind farm power generation revenue.

[0108] This paper conducts experiments based on real-world datasets. The experimental data uses the NREL WIND dataset, which contains multiple variables closely related to wind power, such as wind speed, wind direction, temperature, air pressure, air density, and wind power itself. Historical wind power data and NWP data from 2011 serve as the training set, while historical data from 2012 serves as the test set. The experimental environment uses an Intel® Core™ i5-1035G1 CPU at 1.00GHz and 1.19GHz.

[0109] The parameters of the adaptive prediction model DDA-LightGBM proposed in this invention include the offline LightGBM model initial parameters θ L and DDA algorithm parameters Θ D .

[0110] Offline LightGBM model parameters such as the number of iterations (θ1), the maximum depth of the decision tree (θ2), the number of leaf nodes in the decision tree (θ3), the minimum number of leaf node samples (θ4), the maximum number of leaf node bins (θ5), feature score (θ6), bagging score (θ7), regularization parameter and (θ8,θ9), the minimum gain of node splitting (θ 10 ). The optimal values ​​of these parameters vary in different weather modes.

[0111] The DDA algorithm parameters include Tp1, Tp2, Size (W) and t max Parameter Θ L Optimization using grid search with cross validation, Θ D The tree-structured Parzen estimator (TPE) algorithm is used for optimization, and its value range also varies depending on the meteorological pattern.

[0112] The optimized parameters of the meteorological model based on k=8 are shown in Table 2 below.

[0113] Table 2

[0114]

[0115]

[0116] In this paper, mean absolute error (MAE) and root mean square error (RMSE) are selected as evaluation indicators. The definitions of these indicators are as follows:

[0117]

[0118] Where, Represents wind power y i The smaller the MAE and RMSE values ​​are, the smaller the gap between the power prediction value and the true value is, and the better the prediction effect is.

[0119] In order to verify the effectiveness of the proposed adaptive prediction model for wind turbine power generation based on non-stationary meteorological data, the results of the adaptive prediction model in the sample-sequence feature space (V 2S ), only consider the sample feature space (V sa ) and only consider the sequence feature space (V se ) MAE and RMSE of the prediction results 5 minutes in advance for 9 prediction methods (SVR, RF, Adaboost, GBRT, Bagging, ExtraTree, XGBoost, LightGBM, DDA-LightGBM) in three different feature spaces.

[0120] The evaluation index values ​​of each prediction model are shown in Table 3:

[0121] Table 3

[0122]

[0123] The results show that for each prediction model, the proposed sample-sequence feature extraction method can significantly improve the power prediction accuracy.

[0124] In order to illustrate the similarity search process and show the advantages of the proposed sample-sequence feature in characterizing wind power data features, four fragments are randomly selected from the test dataset and similarity search result graphs based on three different feature spaces are drawn, as shown in Figure 2. Figure 3 shown.

[0125] The results show that the sample-sequence feature can consider both the overall level and temporal characteristics, effectively describe the similarity of data segments, and improve the power prediction accuracy.

[0126] In order to verify the robustness of DDA-LightGBM, four time segments (all half a day in length) were randomly selected and analyzed in the feature space V 2S The prediction results of different methods are compared below, such as Figure 4-7 shown.

[0127] The results show that the DDA-LightGBM model can more accurately track the changing trends of actual wind power output data, especially under high fluctuations, and its prediction results are closer to the actual values ​​than other models. However, in certain time periods (such as when unit components fail), all models may overestimate the power generation.

[0128] The present invention considers multi-step prediction and compares the prediction performance of the model within this range. The results are as follows Figure 8 shown.

[0129] The results show that when the lead time is in the range of 5-30 minutes, the RMSE value of DDA-LightGBM is always the smallest, proving its potential application value. The sample-sequence feature extraction method and the ability to automatically update the model enable it to adapt to complex wind power characteristics.

[0130] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above and that the invention can be embodied in other specific forms without departing from the spirit or essential characteristics of the invention. Therefore, the embodiments should be considered in all respects as illustrative and non-restrictive, and the scope of the invention is defined by the appended claims, not the foregoing description, and all variations within the meaning and range of equivalents of the claims are intended to be included therein. Any reference sign in a claim should not be construed as limiting the claim to which it relates.

[0131] In addition, it should be understood that although this specification is described in terms of implementation methods, not every implementation method contains only one independent technical solution. This narrative method of the specification is only for the sake of clarity. Those skilled in the art should regard the specification as a whole. The technical solutions in each embodiment can also be appropriately combined to form other implementation methods that can be understood by those skilled in the art.

Claims

1. A method for adaptively predicting wind turbine power generation under non-stationary data flow, characterized in that: The following steps are involved: S1. Extract features from numerical weather forecast (NWP) data to obtain sample-related features and sequence-related features, and construct a sample-sequence feature vector. S2. Perform similarity search and clustering based on sample-sequence feature vectors to identify different meteorological patterns; S3. Use historical NWP and power data for each meteorological mode to train a LightGBM network offline to obtain a power prediction model. Power prediction is then performed based on real-time NWP meteorological data. Using drift detection and the DDA algorithm, the prediction performance of the power prediction model under each meteorological mode is monitored in real time, and model parameters are adaptively adjusted to achieve adaptive wind power prediction under non-stationary data streams. Using drift detection and the DDA algorithm, the power prediction model is divided into three states: normal, warning, and drift, using the warning threshold Tp1 and the drift threshold Tp2. The details are as follows: The input NWP data is fed into the model through a sliding window, and the decrease in the model's prediction accuracy for two adjacent windows is calculated: ； in, and is the root mean square sum of the model prediction errors of two adjacent windows; when When , the model remains normal. When When , the model enters the drift state; In different states, the model parameters are updated based on the change in prediction accuracy, which enables adaptive adjustment of the model parameters, including: 1) Under normal conditions, if the current LightGBM model's prediction accuracy drop for two adjacent windows satisfies , then it enters the warning state and the cache window starts collecting new samples; 2) In the early warning state, if the prediction accuracy is further reduced, and If the warning state lasts for less than the set time, the model will be retrained with new samples and enter the drift state, while recording the performance benchmark; if the prediction accuracy is further reduced, but If the warning state lasts longer than the set time, it will be considered a false alarm and return to normal state; 3) In the drift state, the cache window continues to collect samples and continuously evaluates the performance of the current model on the newly collected samples. hour, To establish a new concept’s performance benchmark, the model is retrained using samples in the cache window and the cache window is released after the status returns to normal.

2. The method for adaptively predicting wind turbine generator power generation under non-stationary data flow according to claim 1, characterized in that: The S1 comprises the following sub-steps: S11. Use a sliding window to divide the original input NWP data of each meteorological variable into data segments of equal length. For meteorological variable i, the data segment of length T in the t-th time window is , the meteorological variables include wind speed, wind direction sine, wind direction cosine, temperature, air pressure, and air density; S12. Extract 14 sample-related features from the data segment, including the arithmetic mean, maximum value, minimum value, 5th and 95th percentiles, first, second, and third quartiles, number of occurrences of minimum and maximum values, range, variance, approximate unbiased estimate of skewness, and approximate unbiased estimate of kurtosis of meteorological variable i in time window t; S13. Extracting 19 sequence-related features from the data segment, including: Autocorrelation related features: autocorrelation function ACF, partial autocorrelation function PACF, Box-Pierce statistic, Hurst exponent; Volatility-related characteristics: Detrended volatility analysis (DFA), Lomb-Scargle periodogram, level shift, variance change, crossover point, Shannon entropy, Kullback-Leibler divergence; Time series decomposition related features: trend factor, seasonal factor, detrended series correlation, Spikiness intensity, Lumpiness intensity; Time-related features: the time of occurrence of maximum value, minimum value, and maximum KL divergence; S14. Normalize the extracted sample-related features and sequence-related features, and combine the features of the six meteorological variables to form a 198-dimensional sample-sequence feature vector of the data segments divided by the same time window.

3. The method for adaptively predicting wind turbine generator power generation under non-stationary data flow according to claim 2, characterized in that: The trend factors are: ； The seasonal factors are: ； in, is the variance function; is the deseasonal sequence; is a detrended series; is a stable normalized time series after Box-Cox transformation; The trend component, seasonal component and residual component obtained by LOESS decomposition respectively; The detrended correlation is: ； Where, is the autocorrelation function of the detrended NWP series i at the kth lag in the time window t, n is the length of the series, and h is the maximum lag considered.

4. The method for adaptively predicting wind turbine generator power generation under non-stationary data flow according to claim 1, characterized in that: The S2 includes the following sub-steps: S21. Perform similarity search based on the sample-sequence feature vector by calculating the Euclidean distance; S22. The k-means++ algorithm is used to cluster data segments based on similarity search results. The optimal number of meteorological patterns is determined by the elbow rule, and similar data segments are clustered into different meteorological patterns.

5. The method for adaptively predicting wind turbine generator power generation under non-stationary data flow according to claim 1, characterized in that: The S3 includes the following sub-steps: S31. For each meteorological pattern, use historical NWP data and its corresponding historical power data to train the LightGBM network offline to obtain the power prediction model for the corresponding meteorological pattern; S32. Real-time NWP data collection, extraction of sample-sequence feature vectors, calculation of the similarity between the sample-sequence feature vectors and various meteorological patterns, matching the corresponding meteorological patterns, and inputting the real-time NWP data into the power prediction model corresponding to the meteorological pattern for power prediction; S33. Using drift detection and DDA algorithms, the power prediction model is divided into three states: normal, warning, and drift through the warning threshold Tp1 and drift threshold Tp2. The model prediction performance is monitored in real time. In different states, the decision on whether to update the model parameters is made based on the change in prediction accuracy, thus achieving adaptive adjustment of the model parameters.

Images