A wind farm power probability prediction method based on low rank adaptation

Abstract

The application discloses a wind farm power probability prediction method based on low rank adaptation, which comprises the following steps: collecting operation data of a wind farm in a period of time, determining characteristic parameter values and pre-processing; then, performing prior enhancement processing on all characteristic parameters based on physical mechanism, and constructing training samples based on the enhanced characteristic parameters; then, establishing a time series prediction model for predicting the time-varying condition of the output active power of the wind farm; and finally, predicting the active power of the wind farm at a future time through the time series prediction model, combining the early operation state and weather condition of the wind farm, and outputting the prediction result. The method has high prediction accuracy through the prior enhancement processing of the physical mechanism of the input data, can output the prediction value and the prediction interval at the same time, and meets the demand of the operation state risk assessment; the time series prediction model adopts a low rank adaptation mechanism, has strong generalization ability, and is beneficial to rapid deployment.

Description

Technical Field

[0001] This invention belongs to the field of low-rank adaptation and wind power prediction technology, specifically relating to a wind farm power probability prediction method based on low-rank adaptation. Background Technology

[0002] With the rapid development of the wind power industry, the proportion of wind power installed capacity in the total installed capacity of the power system is constantly increasing. Due to the obvious volatility, intermittency, and randomness of wind energy, the output power of wind turbines is significantly unstable. Therefore, wind power grid connection can easily lead to frequency and voltage fluctuations in the power grid, affecting the stability and security of the grid. One way to solve the above problems is to predict wind power to regulate the power grid. For such methods, the accuracy of power prediction directly affects the grid connection effect. With the rapid development of neural networks in recent years, various deep learning models based on neural networks have gradually begun to be applied to the field of wind power prediction. For example, Chinese patent application, titled "A Spatiotemporal Prediction Method and System for Wind Power Based on Graph Neural Network and Dilated Temporal Convolution", application number: 202511636857.7, proposes a model based on graph neural network with wind turbine location as the node. This model uses raw wind farm data, including wind speed, wind direction, and nacelle temperature, as input to derive the predicted power value. Methods that directly apply deep learning models typically suffer from three main problems: First, they directly use raw wind farm data such as wind speed and direction as input, lacking explicit utilization of the wind power conversion mechanism. This makes it difficult to accurately characterize the nonlinear relationship between wind speed and power, as well as the wake effect caused by wind direction changes, and can easily lead to large errors when wind power changes significantly. Second, the prediction result is a single power value, which does not contain uncertainty information and cannot meet the needs of wind farm operation status risk assessment. Third, existing neural network methods are all built for a specific wind farm. When used for other wind farms, they need to be retrained and modeled, thus requiring high computational and storage resources, lacking transfer adaptability, and hindering rapid deployment.

[0003] Several neural network-based methods have been proposed to address the aforementioned issues. For example, Chinese patent application No. 202512010710.3, titled "A Wind Power Prediction Method and System Based on Deep Learning," proposes a model based on Long Short-Term Memory (LSTM) neural networks that learns the temporal characteristics of raw wind farm data. This model outputs a probabilistic prediction interval for wind power. Another example is Chinese patent application No. 202511903121.1, titled "Wind Power Prediction System and Method Integrating Signal Decomposition and Deep Learning," which proposes a method to decompose meteorological and wind power data into multiple subsequences through two-level signal decomposition. Each subsequence is then predicted using an LSTM-based model, and all predictions are fused to obtain the final power value. It can be seen that these two methods only improve upon certain shortcomings of neural network-based methods, without significantly improving overall performance, and neither solves the problem of rapid deployment.

[0004] In summary, this paper proposes a neural network-based wind power prediction method that can analyze the wind power and wind power conversion mechanism based on original wind farm data, provide information on the uncertainty of prediction results, and has strong transfer adaptability. This method is of great significance for promoting wind power grid connection, ensuring grid dispatch and operation safety, and improving wind power absorption efficiency. Summary of the Invention

[0005] The purpose of this invention is to address the problems of existing neural network-based wind power prediction methods, such as lack of analysis of wind power conversion mechanisms, prediction results consisting only of power values, and lack of transfer adaptation capabilities. This invention proposes a wind farm power probability prediction method based on low-rank adaptation by performing physical mechanism prior enhancement processing on preprocessed wind farm data and then establishing a prediction model through a low-rank adaptation mechanism.

[0006] The technical solution adopted in this invention is as follows:

[0007] A wind farm power probability prediction method based on low-rank adaptation specifically includes the following steps:

[0008] S1. Collect the operating data of the wind farm that needs to be predicted for a period of time, including unit power data and meteorological data. At the same time, set a variety of time-varying characteristic parameters, determine the value of the characteristic parameters based on the wind farm operating data, and perform preprocessing.

[0009] S2. Perform prior augmentation on all feature parameters based on physical mechanisms, and construct training samples based on the augmented feature parameters.

[0010] S3. Establish a time-series prediction model for predicting the change of active power output at wind farms over time.

[0011] S4. Using the aforementioned time-series prediction model, combined with the wind farm's earlier operating status and meteorological conditions, predict the active power output of the wind farm at future times, and output the prediction results.

[0012] To address the shortcomings of existing neural network-based wind farm power prediction methods, such as difficulty in characterizing the nonlinear relationship between meteorological conditions and power, and high deployment costs across wind farms, this invention proposes a power prediction method that integrates physical prior enhancement and low-rank adaptation. The physical quantities enhanced by prior enhancement are beneficial for explicitly characterizing the nonlinear coupling relationship between wind speed and output power, thereby improving the accuracy of wind power prediction. The low-rank adaptation technique enables training only on low-rank adaptation parameters while keeping the parameters in the main model frozen, thus achieving rapid adaptation to the target wind farm, enhancing the model's generalization ability in cross-wind farm scenarios, and reducing model deployment and update costs.

[0013] Further optimization, step S1 specifically includes the following steps:

[0014] S1.1. Multiple sampling stages, sampling intervals, and multiple characteristic parameters are set. All sampling stages have equal durations. Multiple consecutive sampling times are determined within each sampling stage according to the sampling intervals. Simultaneously, unit power data and meteorological data corresponding to each sampling stage are collected. Based on the above data, the values ​​of various characteristic parameters at each sampling time within the sampling stage are determined. The characteristic parameters include active power P. t Wind speed v t Wind direction θ t and ambient temperature T t .

[0015] S1.2. Set the missing threshold N t If no feature parameter values ​​are missing within a certain sampling stage, then that sampling stage is marked as valid. If a feature parameter exists within a certain sampling stage that is consecutively N... t If all values ​​are missing within a given time period, that sampling phase is marked as invalid. If a feature parameter value is missing within a certain sampling phase, and the number of consecutive missing values ​​at any given time does not reach N, then the sampling phase is marked as invalid. t If so, then the sampling phase is marked as fillable.

[0016] S1.3. Based on wind direction θ t The sinusoidal component s is obtained t With cosine component c t s t = sin(θ t ), c t = cos(θ t For each marker-fillable sampling stage, missing feature parameter values ​​are filled in. The type of the filled values ​​includes P.t T t v t s t c t .

[0017] Further optimization involves the following steps in step S1.3: Filling in the missing feature parameter values.

[0018] S1.3.1. Record the time corresponding to the missing feature parameter values ​​as t. a In the sampling phase where the missing location is located, if t a If only one side of the past and future sides has a feature parameter value of the same type as the missing value, then the distance t will be... a Let t1 be the most recent time when a value of the same type exists, and let x be the value of the same type at time t1. t1 , used to fill t a The value at the missing time point is denoted as x. ta Then x ta = x t1 .

[0019] S1.3.2. In the sampling phase where the missing location is located, if t a If both the past and future sides of a value contain feature parameters of the same type as the missing value, then the distance t on the past side will be... a The most recent moment with the same value is denoted as t1, and the future distance to one side is t. a Let t2 be the most recent time when the same value exists, and let x be the same value at times t1 and t2 respectively. t1 x t2 Then x can be obtained through the following formula. ta :

[0020] (1)

[0021] For any time t where the value of a characteristic parameter is missing a First, it determines the existence of similar feature parameters on both the past and future sides. Then, it selects the appropriate process from the two processes mentioned above to obtain the corresponding x. ta .

[0022] Considering that the wind farm field data acquisition system may be affected by factors such as sensor fluctuations, communication interruptions or storage anomalies during long-term operation, there may be local missing data in the original data. In order to ensure the quality of subsequent training, the missing parts need to be filled before constructing training samples. Linear interpolation is used to fill the missing parts when there are valid values ​​on both the past and future sides, and boundary filling is used when there are valid values ​​on only one side.

[0023] Further optimization, step S2 specifically includes the following steps:

[0024] S2.1. Wind speed v at time t t The cubic eigenvalue v of the wind speed at that moment is obtained. h t v h t = v t 3 .

[0025] S2.2. Taking each sampling stage marked as valid and each marked as fillable with all missing values ​​filled as a sample stage, construct the feature vector for all sampling times in each sample stage. The feature vector for any sampling time t is represented as S t = [P t , T t , v t , v h t , s t , c t ].

[0026] S2.3. Set a prediction stage corresponding to each sample stage, wherein the initial time of the prediction stage is the last time of the corresponding sample stage, and the duration of all prediction stages is equal; determine multiple consecutive prediction times within the prediction stage according to the sampling interval, and collect meteorological data of all prediction stages to determine v at each prediction time. t θ t And T t And construct a covariate vector, the covariate vector at any prediction time t is represented as V t = [T t , v t , v h t , s t , c t ].

[0027] S2.4. Establish a time-series basic model, which has a trainable low-rank parameter branch. The model takes the feature vector of the sample stage and the covariate vector of the corresponding prediction stage as inputs, and outputs the active power probability distribution at each prediction time in the corresponding prediction stage. Collect wind turbine power data and meteorological data under the same scenario, and pre-train the main body of the time-series basic model using the collected data.

[0028] S2.5. Select a sample stage as the test sample stage, and the prediction stage corresponding to this sample stage is the final prediction stage; arbitrarily select a prediction stage other than the final prediction stage, input all covariate vectors and all feature vectors of the sample stage corresponding to this prediction stage into the pre-trained time series basic model, obtain the active power probability distribution of all prediction times in the corresponding prediction stage, and then determine the predicted values ​​for all prediction times. Then, take another prediction stage that has not yet obtained a predicted value and repeat the above process until all prediction stages except the final prediction stage have obtained corresponding predicted values.

[0029] S2.6. Add each predicted active power value to the covariate vector at the corresponding prediction time to obtain the feature vector at the prediction time; construct multiple training samples, wherein the number of training samples is equal to and corresponds one-to-one with all sample stages except the test sample stage, and the training samples are a set containing all feature vectors of the corresponding sample stage and the prediction stage.

[0030] By introducing the cubic eigenvalue of wind speed v h t This invention can explicitly express the nonlinear relationship between wind speed and power during wind power conversion, thereby reducing the learning burden on the model during training. The prior augmentation process in this invention includes constructing the cubic eigenvalues ​​of wind speed and the sine and cosine components of wind direction. Training the model with the data after prior augmentation is beneficial to improving the model's response to changes in meteorological conditions.

[0031] Further optimization, step S3 specifically includes the following steps:

[0032] S3.1. Set a test sample, which is a set containing all feature vectors of the test sample stage and all covariate vectors of the final prediction stage. At the same time, collect the power data of the wind turbine in the final prediction stage, and derive the true value of active power at all prediction times in the final prediction stage based on the collected data.

[0033] S3.2. Set the round threshold and discrimination conditions. Perform multiple rounds of adaptation training on the pre-trained time series basic model using the training samples. During training, only the parameters in the low-rank parameter branches are updated, while all parameters in the main body of the model remain unchanged. After the training round reaches the round threshold, use the obtained model combined with the test samples to predict the active power at all prediction times in the final prediction stage. Based on the probability distribution of active power, obtain the predicted values ​​for all prediction times. Then, compare the actual active power values ​​with the predicted values ​​and determine whether the comparison results meet the discrimination conditions. If the discrimination conditions are met, the adaptation training is considered complete, and the obtained model is the time series prediction model. If not, adjust the training samples and continue training the model according to the above process until the discrimination conditions are met.

[0034] Considering the significant differences in terrain conditions, turbulence intensity, turbine status, and seasonal distribution among different wind farms, directly using a general model, while possessing some predictive ability, is insufficient to fully reflect the local characteristics of the target wind farm. Therefore, this invention employs a low-rank adaptation mechanism. During the adaptation training process, the backbone of the model, already pre-trained on large-scale multi-source time series, remains unchanged; only the low-rank parameter branches are trained. Unlike traditional models trained from scratch, this method eliminates the need to rebuild all parameters from scratch when the wind farm changes. Instead, it allows for targeted adaptation based on existing time series knowledge, significantly improving training efficiency and model generalization ability.

[0035] Further optimization is needed; the specific discrimination conditions in step S3.2 are as follows:

[0036] Four metrics are defined: Mean Absolute Error (MAE), Root Mean Square Error (RMSE), Interval Coverage (PICP), and Mean Interval Width (MPIW), along with their corresponding first, second, third, and fourth thresholds. These four metrics are calculated using the following formulas:

[0037] (2)

[0038] Where MAE, RMSE, PICP, and MPIW represent the mean squared error, mean absolute error, interval coverage, and average interval width, respectively, and N is the number of prediction times included in the final prediction stage. In this embodiment, N = 48. t y' t Let It be the predicted value and the actual value of active power at time t, respectively, and let II(·) be the indicator function. t L t These are the upper and lower bounds of the power value distribution interval obtained from the probability distribution of active power, respectively.

[0039] The obtained MAE, RMSE, PICP, and MPIW are compared with the first, second, third, and fourth thresholds, respectively. If the MAE does not exceed the first threshold, the RMSE does not exceed the second threshold, the PICP is not lower than the third threshold, and the MPIW does not exceed the fourth threshold, then the discrimination condition is considered to be met; otherwise, the discrimination condition is considered not met.

[0040] Further optimization, step S4 specifically includes the following steps:

[0041] S4.1. Set the sampling stage and prediction stage with the current time as the reference time. Construct the feature vector of the sampling stage based on the unit power data and meteorological data of the past time. Construct the covariate vector of the prediction stage based on the meteorological forecast. Input the constructed feature vector and covariate vector into the time series prediction model to obtain the probability distribution of all prediction times.

[0042] S4.2. Establish a quantile prediction model. This model is used to determine the distribution interval of a value based on the probability distribution, and, combined with the set numerical distribution conditions, to obtain the position of a certain probability value within the distribution interval. This position is called the probability quantile. The quantile prediction model is described by the following formula:

[0043] (3)

[0044] Where P(·) is the probability function, q is the probability value, and y is the probability value. (q) t Let q be the probability quantile corresponding to the probability value q based on the probability distribution at time t.

[0045] S4.3. Set three probability values: lower limit, median, and upper limit, denoted as q. l q m q u , q l q m q u Substituting these values ​​into the quantile prediction model yields the corresponding probability quantile y. (ql) t y (qm) t y (qu) t , will y (ql) t and y (qu) t Using these as the upper and lower limits of the interval respectively, the quantile prediction interval [y] is obtained. (ql) t , y (qu) t ], y (qm) t As the predicted active power value at prediction time t, [y (ql) t , y (qu) t The active power prediction interval for prediction time t is used as the final output, which is the y value corresponding to all prediction times. (qm) t With [y (ql) t , y (qu) t ].

[0046] The above quantile output method can not only provide power prediction values, but also the range of likely future power fluctuations. When the prediction range is narrow, it indicates that the future wind conditions are relatively stable. When the prediction range is significantly expanded, it indicates that the future power fluctuations are highly uncertain, and a greater safety margin needs to be left in energy storage scheduling and reserve capacity configuration.

[0047] The beneficial effects of the method of the present invention are as follows:

[0048] 1. This invention constructs physical enhancement inputs by using the cubic features of wind speed and the sine and cosine components of wind direction, enabling the pre-trained time series basic model to explicitly utilize the wind power conversion mechanism during the prediction process, reducing the difficulty for the model to learn the nonlinear relationship between wind speed and power and the periodic relationship of wind direction, and improving the accuracy of power prediction;

[0049] 2. This invention combines physical augmentation input with low-rank adaptation training. By freezing the main parameters of the pre-trained time series basic model and updating only the low-rank adaptation parameters, the model can be adapted to the local operating characteristics of the target wind farm, thereby reducing the cost of cross-wind farm deployment and model update.

[0050] 3. This invention outputs probability quantiles through an adapted time-series prediction model and a quantile prediction model, thereby obtaining point prediction values ​​and prediction intervals. This enables the model to not only provide future power change trends but also characterize the future power fluctuation range, providing a basis for wind power dispatch, reserve capacity configuration, and operational risk assessment. Attached Figure Description

[0051] Figure 1 A schematic diagram of the overall process of the wind power prediction method of this invention.

[0052] Figure 2 A comparison chart of the prediction effects of the method of the present invention under different application levels.

[0053] Figure 3 Comparison of prediction intervals and actual values ​​for different confidence levels.

[0054] Figure 4 The prediction results of the method of this invention are compared with those of other methods and the time-varying curves of the actual values. Detailed Implementation

[0055] To make the objectives, technical solutions, and advantages of the present invention clearer, the technical solutions of the present invention will be clearly and completely described below through specific embodiments. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0056] Example 1: A wind farm power probability prediction method based on low-rank adaptation, specifically including the following steps:

[0057] S1. Collect operational data of the wind farms for which power prediction is required over a period of time, including turbine power data and meteorological data, and preprocess the obtained wind farm operational data, specifically including the following steps:

[0058] S1.1. Multiple sampling stages, sampling intervals, and multiple characteristic parameters are set. All sampling stages have equal durations. Multiple consecutive sampling times are determined within each sampling stage according to the sampling intervals. Simultaneously, unit power data and meteorological data corresponding to each sampling stage are collected. The values ​​of various characteristic parameters at each sampling time within the sampling stage are determined based on the aforementioned data. In this embodiment, multiple reference times are first set over a period of time, with an interval of 2 hours between any two adjacent reference times. The number of reference times is equal to and corresponds one-to-one with the number of sampling stages. Sampling stages are then set according to the reference times, with each reference time being the last time of its corresponding sampling stage. The sampling interval is 10 minutes, and each sampling stage contains 512 sampling times. The characteristic parameter is active power P. t Wind speed v t Wind direction θ t and ambient temperature T t Four terms, where t is time, P t The value is based on the average active power P' of all wind turbines in the wind farm. t Determine if P' is obtained at a certain moment t If the value is non-negative, then P at that moment... t = P' t If P' is obtained at a certain moment t If the value is negative, then at that moment P t = 0.

[0059] S1.2. Set the missing threshold N t If no feature parameter values ​​are missing within a certain sampling stage, then that sampling stage is marked as valid. If a feature parameter exists within a certain sampling stage that is consecutively N... t If all values ​​are missing within a given time period, that sampling phase is marked as invalid. If a feature parameter value is missing within a certain sampling phase, and the number of consecutive missing values ​​at any given time does not reach N, then the sampling phase is marked as invalid. t If so, then the sampling phase is marked as fillable.

[0060] S1.3. Based on wind direction θ t The sinusoidal component s is obtained t With cosine component c t s t = sin(θ t), c t = cos(θ t For each marker-fillable sampling stage, missing feature parameter values ​​are filled in. The type of the filled values ​​includes P. t T t v t s t c t The specific filling method is as follows:

[0061] S1.3.1. Record the time corresponding to the missing feature parameter values ​​as t. a In the sampling phase where the missing location is located, if t a If only one side of the past and future sides has a feature parameter value of the same type as the missing value, then the distance t will be... a Let t1 be the most recent time when a value of the same type exists, and let x be the value of the same type at time t1. t1 , used to fill t a The value at the missing time point is denoted as x. ta Then x ta = x t1 .

[0062] S1.3.2. In the sampling phase where the missing location is located, if t a If both the past and future sides of a value contain feature parameters of the same type as the missing value, then the distance t on the past side will be... a The most recent moment with the same value is denoted as t1, and the future distance to one side is t. a Let t2 be the most recent time when the same value exists, and let x be the same value at times t1 and t2 respectively. t1 x t2 And x is obtained through formula (1) ta .

[0063] For any time t where the value of a characteristic parameter is missing a First, it determines the existence of similar feature parameters on both the past and future sides. Then, it selects the appropriate process from the two processes mentioned above to obtain the corresponding x. ta .

[0064] S2. Perform prior enhancement processing on all feature parameters based on physical mechanisms, and construct training samples based on the processed feature parameters. This includes the following steps:

[0065] S2.1. Wind speed v at time t t The cubic eigenvalue v of the wind speed at that moment is obtained. h t v h t = v t3 .

[0066] S2.2. Taking each sampling stage marked as valid and each marked as fillable with all missing values ​​filled as a sample stage, construct the feature vector for all sampling times in each sample stage. The feature vector for any sampling time t is represented as S t = [P t , T t , v t , v h t , s t , c t ].

[0067] S2.3. A prediction stage is defined for each sample stage, with the initial time of the prediction stage being the last time of the corresponding sample stage, and all prediction stages having equal duration. Multiple consecutive prediction times are determined within the prediction stage based on the sampling interval. In this embodiment, the prediction stage includes 48 prediction times. Meteorological data from all prediction stages are collected to determine the v of each prediction time. t θ t And T t And construct a covariate vector, the covariate vector at any prediction time t is represented as V t = [T t , v t , v h t , s t , c t ].

[0068] S2.4. Establish a time-series basic model. The attention projection part of the time-series basic model has a trainable low-rank parameter branch. The model takes the feature vector of the sample stage and the covariate vector of the corresponding prediction stage as inputs, and the output is the probability distribution of active power at each prediction time in the corresponding prediction stage. In this embodiment, the time-series basic model is the chronos-2 model. Collect wind turbine power data and meteorological data in the same scenario, and pre-train the main body of the time-series basic model with the collected data so that the main body of the model has the ability to predict the basic time-varying characteristics of wind turbine active power under the influence of weather factors.

[0069] S2.5. Select a sample stage as the test sample stage, and the prediction stage corresponding to this sample stage is the final prediction stage; arbitrarily select a prediction stage other than the final prediction stage, input all covariate vectors and all feature vectors of the sample stage corresponding to this prediction stage into the pre-trained time series basic model, obtain the active power probability distribution of all prediction times in the corresponding prediction stage, and then determine the predicted values ​​for all prediction times. Then, take another prediction stage that has not yet obtained a predicted value and repeat the above process until all prediction stages except the final prediction stage have obtained corresponding predicted values.

[0070] S2.6. Add each predicted active power value to the covariate vector at the corresponding prediction time to obtain the feature vector at the prediction time; construct multiple training samples, wherein the number of training samples is equal to and corresponds one-to-one with the number of all sample stages except the test sample stage, and the training samples are a set containing all feature vectors of the corresponding sample stage and the prediction stage. In this way, a new training sample can be generated every 2 hours within a certain period of time. On the one hand, this can significantly increase the number of samples and improve the stability of model training; on the other hand, it can still retain a certain difference between adjacent samples, so that the model can repeatedly learn the evolution pattern of active power in different weather stages and different power change ranges.

[0071] Due to wind direction θ t The angle value changes periodically. If the angle value is directly interpolated, it is easy to cause large numerical errors near 0° and 360°. Therefore, the wind direction is converted into sine and cosine components.

[0072] S3. Establish a time-series prediction model for predicting the change of active power output at wind farms over time, as detailed below:

[0073] S3.1. Set a test sample, which is a set containing all feature vectors of the test sample stage and all covariate vectors of the final prediction stage. At the same time, collect the power data of the wind turbine in the final prediction stage, and derive the true value of active power at all prediction times in the final prediction stage based on the collected data.

[0074] S3.2. Set the round threshold and discrimination conditions, and perform multiple rounds of adaptation training on the pre-trained time series basic model using the training samples. In this embodiment, adaptation training is performed using Low-Rank Adaptation (LoRA). During training, only the parameters in the low-rank parameter branches are updated, while all parameters in the main body of the model remain unchanged. In this embodiment, the training steps are 1000 steps, and the learning rate is 1×10⁻⁶. -4The batch size is 16. After the training rounds reach the threshold, the obtained model is combined with the test samples to predict the active power at all prediction times in the final prediction stage. The predicted values ​​for all prediction times are obtained based on the probability distribution of active power. Then, the actual active power values ​​are compared with the predicted values, and it is determined whether the comparison results meet the discrimination conditions. If the discrimination conditions are met, the adaptation training is considered to be completed, and the obtained model is the time series prediction model. If the conditions are not met, the training samples are adjusted, and the model is trained again according to the above process until the discrimination conditions are met.

[0075] In this embodiment, four indicators are set: mean squared error (MSE), mean absolute error (MAE), interval coverage (PICP), and mean interval width (MPIW), as well as the first, second, third, and fourth thresholds corresponding to the four indicators. The above four indicators are calculated by formula (2). In this embodiment, N = 48.

[0076] The obtained MAE, RMSE, PICP, and MPIW are compared with the first, second, third, and fourth thresholds, respectively. If the MAE does not exceed the first threshold, the RMSE does not exceed the second threshold, the PICP is not lower than the third threshold, and the MPIW does not exceed the fourth threshold, then the discrimination condition is considered to be met; otherwise, the discrimination condition is considered not met.

[0077] S4. Using the aforementioned time-series prediction model, combined with the wind farm's earlier operating status and meteorological conditions, the active power output of the wind farm at future times is predicted, and the prediction results are output. The specific process is as follows:

[0078] S4.1. Set the sampling stage and prediction stage using the current time as the reference time. Construct the feature vector of the sampling stage based on the unit power data and meteorological data of the past time. Construct the covariate vector of the prediction stage based on the meteorological forecast. Input the constructed feature vector and covariate vector into the time series prediction model to obtain the probability distribution of all prediction times.

[0079] S4.2. Establish a quantile prediction model. This model is used to determine the distribution interval of the value based on the probability distribution and, in combination with the set value distribution conditions, to obtain the position of a certain probability value in the distribution interval. This position is called the probability quantile point. The quantile prediction model is described by formula (3).

[0080] S4.3. Set three probability values: lower limit, median, and upper limit, denoted as q. l q m q u , q l q m q u Substituting these values ​​into the quantile prediction model yields the corresponding probability quantile y.(ql) t y (qm) t y (qu) t , will y (ql) t and y (qu) t Using these as the upper and lower limits of the interval respectively, the quantile prediction interval [y] is obtained. (ql) t , y (qu) t ], y (qm) t As the predicted active power value at prediction time t, [y (ql) t , y (qu) t The active power prediction interval for prediction time t is used as the final output, which is the y value corresponding to all prediction times. (qm) t With [y (ql) t , y (qu) t In this embodiment, q l = 0.1, q m = 0.5, q u = 0.9, prediction interval confidence level 80%, that is, the actual active power value at time t falls within the quantile prediction interval [y (0.1) t , y (0.9) t The probability in the figure is 80%, and the output is the predicted active power value and prediction range for the next 8 hours. The overall process of the method described in this invention is as follows: Figure 1 As shown.

[0081] The predictive effect of the method of the present invention under different application levels is as follows: Figure 2 As shown, where Figure 2 (a) Figure 2 (b) Figure 2 (c) The prediction results are obtained by inputting only the feature vector at the sampling time, inputting the feature vector at the sampling time and the co-meteorological data at the prediction time, and inputting the feature vector at the sampling time and the covariate vector at the prediction time, respectively. Figure 2 (d) shows the prediction results obtained through the complete process of the method of this invention. The horizontal axis represents time, and the vertical axis represents active power values ​​(unit: kW). Comparing the four figures, it can be seen that from... Figure 2 (a) to Figure 2(d) The similarity between the predicted and actual value curves gradually increases, while the predicted interval gradually decreases as the overlap with the actual value curve remains relatively stable. The prediction results obtained by the method of this invention are as follows: Figure 3 As shown in the figure, the horizontal axis represents time, and the vertical axis represents the active power value (unit: kW). In the legend, CI is an abbreviation for confidence interval. It can be seen from the figure that the predicted value obtained by the method of this invention has roughly the same trend as the actual value over time. By comparing the range of the prediction interval obtained at three confidence levels of 50%, 80%, and 90%, it can be seen that when the confidence level reaches 80%, the actual value can basically fall within the prediction interval.

[0082] The effects of the method of the present invention compared with other existing methods are as follows: Figure 4 As shown in the figure, the horizontal axis represents time, and the vertical axis represents the active power value (unit: kW). The method of this invention provides the predicted value and the prediction interval, with the prediction interval set as [y (0.1) t , y (0.9) t Other methods only provide predicted values. The left half of the figure represents the sampling stage, and the right half represents the prediction stage. The prediction stage includes prediction results obtained using the method of this invention, as well as prediction results obtained using three methods: LightGBM, PatchTST, and LSTM. Based on the comparison between the predicted values ​​and the actual values ​​for each method in the prediction stage, the mean absolute error (MAE) of this invention is 56.23, and the root mean square error (RMSE) is 77.12. The MAE of the LightGBM method is 197.0, and the RMSE is 262.77. The MAE of the PatchTST method is 151.19. The RMSE of the method is 215.76, while the MAE of the LSTM method is 132.0 and the RMSE is 197.91. Therefore, the error between the predicted value and the true value obtained by the method of this invention is significantly smaller than that of other methods. At the same time, it can be seen from the figure that the time-varying curve of the predicted value obtained by the method of this invention has a higher degree of overlap with the time-varying curve of the true value. The curve obtained by the LSTM method has an overly flat trend, while the curves obtained by the LightGBM and PatchTST methods have significant deviations from the curves of the true value. Combining the MAE and RMSE data, as well as the degree of overlap between the predicted value and the time-varying curve of the true value, it can be seen that the method of this invention has higher accuracy.

Claims

1. A wind farm power probability prediction method based on low-rank adaptation, characterized in that, Specifically, the following steps are included: S1. Collect the operating data of the wind farm that needs to be predicted for a period of time, including the unit power data and meteorological data. At the same time, set a variety of time-varying characteristic parameters, determine the value of the characteristic parameters based on the wind farm operating data, and perform preprocessing. S2. Perform prior augmentation on all feature parameters based on physical mechanisms, and construct training samples based on the augmented feature parameters; S3. Establish a time-series prediction model for predicting the change of active power output at wind farms over time; S4. Using the aforementioned time-series prediction model, combined with the wind farm's earlier operating status and meteorological conditions, predict the active power output of the wind farm at future times, and output the prediction results.

2. The wind farm power probability prediction method based on low-rank adaptation as described in claim 1, characterized in that, Step S1 specifically includes the following steps: S1.

1. Multiple sampling stages, sampling intervals, and multiple characteristic parameters are set. All sampling stages have equal durations. Multiple consecutive sampling times are determined within each sampling stage according to the sampling intervals. Simultaneously, unit power data and meteorological data corresponding to each sampling stage are collected. Based on the above data, the values ​​of various characteristic parameters at each sampling time within the sampling stage are determined. The characteristic parameters include active power P. t Wind speed v t Wind direction θ t and ambient temperature T t ; S1.

2. Set the missing threshold N t If no feature parameter values ​​are missing within a certain sampling stage, then that sampling stage is marked as valid. If a feature parameter exists within a certain sampling stage that is consecutively N... t If all values ​​are missing within a given time period, that sampling phase is marked as invalid. If a feature parameter value is missing within a certain sampling phase, and the number of consecutive missing values ​​at any given time does not reach N, then the sampling phase is marked as invalid. t If so, then mark the sampling stage as fillable; S1.

3. Based on wind direction θ t The sinusoidal component s is obtained t With cosine component c t s t = sin(θ t ), c t = cos(θ t For each marker-fillable sampling stage, missing feature parameter values ​​are filled in. The type of the filled values ​​includes P. t T t v t s t c t .

3. The wind farm power probability prediction method based on low-rank adaptation as described in claim 2, characterized in that, The specific process for filling in the missing feature parameter values ​​in step S1.3 includes the following steps: S1.3.

1. Record the time corresponding to the missing feature parameter values ​​as t. a In the sampling phase where the missing location is located, if t a If only one side of the past and future sides has a feature parameter value of the same type as the missing value, then the distance t will be... a Let t1 be the most recent time when a value of the same type exists, and let x be the value of the same type at time t1. t1 , used to fill t a The value at the missing time point is denoted as x. ta Then x ta = x t1 ; S1.3.

2. In the sampling phase where the missing location is located, if t a If both the past and future sides of a value contain feature parameters of the same type as the missing value, then the distance t on the past side will be... a The most recent moment with the same value is denoted as t1, and the future distance to one side is t. a Let t2 be the most recent time when the same value exists, and let x be the same value at times t1 and t2 respectively. t1 x t2 Then x can be obtained through the following formula. ta : (1) For any time t where the value of a characteristic parameter is missing a First, it determines the existence of similar feature parameters on both the past and future sides. Then, it selects the appropriate process from the two processes mentioned above to obtain the corresponding x. ta .

4. The wind farm power probability prediction method based on low-rank adaptation as described in claim 1, characterized in that, Step S2 specifically includes the following steps: S2.

1. Wind speed v at time t t The cubic eigenvalue v of the wind speed at that moment is obtained. h t v h t = v t 3 ; S2.

2. Taking each sampling stage marked as valid and each marked as fillable with all missing values ​​filled as a sample stage, construct the feature vector for all sampling times in each sample stage. The feature vector for any sampling time t is represented as S t = [P t , T t , v t , v h t , s t , c t ]; S2.

3. Set a prediction stage corresponding to each sample stage, wherein the initial time of the prediction stage is the last time of the corresponding sample stage, and the duration of all prediction stages is equal; determine multiple consecutive prediction times within the prediction stage according to the sampling interval, and collect meteorological data of all prediction stages to determine v at each prediction time. t θ t And T t And construct a covariate vector, the covariate vector at any prediction time t is represented as V t = [T t , v t , v h t , s t , c t ]; S2.

4. Establish a time-series basic model, which has a trainable low-rank parameter branch. The model takes the feature vector of the sample stage and the covariate vector of the corresponding prediction stage as inputs, and outputs the active power probability distribution at each prediction time in the corresponding prediction stage. Collect wind turbine power data and meteorological data in the same scenario, and pre-train the main body of the time-series basic model using the collected data. S2.

5. Select a sample stage as the test sample stage, and the prediction stage corresponding to this sample stage is the final prediction stage; take any prediction stage other than the final prediction stage, input all covariate vectors and all feature vectors of the sample stage corresponding to this prediction stage into the pre-trained time series basic model, obtain the active power probability distribution of all prediction times in the corresponding prediction stage, and then determine the predicted value of all prediction times. Then take another prediction stage that has not yet obtained a predicted value and repeat the above process until all prediction stages except the final prediction stage have obtained the corresponding predicted value. S2.

6. Add each predicted active power value to the covariate vector at the corresponding prediction time to obtain the feature vector at the prediction time; construct multiple training samples, wherein the number of training samples is equal to and corresponds one-to-one with all sample stages except the test sample stage, and the training samples are a set containing all feature vectors of the corresponding sample stage and the prediction stage.

5. The wind farm power probability prediction method based on low-rank adaptation as described in claim 1, characterized in that, Step S3 specifically includes the following steps: S3.

1. Set a test sample, which is a set containing all feature vectors of the test sample stage and all covariate vectors of the final prediction stage. At the same time, collect the power data of the wind turbine in the final prediction stage, and derive the true value of active power at all prediction times in the final prediction stage based on the collected data. S3.

2. Set the round threshold and discrimination conditions. Perform multiple rounds of adaptation training on the pre-trained time series basic model using the training samples. During training, only the parameters in the low-rank parameter branches are updated, while all parameters in the main body of the model remain unchanged. After the training round reaches the round threshold, use the obtained model combined with the test samples to predict the active power at all prediction times in the final prediction stage. Based on the probability distribution of active power, obtain the predicted values ​​for all prediction times. Then, compare the actual active power values ​​with the predicted values ​​and determine whether the comparison results meet the discrimination conditions. If the discrimination conditions are met, the adaptation training is considered complete, and the obtained model is the time series prediction model. If not, adjust the training samples and continue training the model according to the above process until the discrimination conditions are met.

6. The wind farm power probability prediction method based on low-rank adaptation as described in claim 5, characterized in that, The specific discrimination conditions described in step S3.2 are as follows: Four indicators are defined: mean absolute error, root mean square error, interval coverage, and average interval width, along with their corresponding first, second, third, and fourth thresholds. These four indicators are calculated using the following formulas: (2) Where MAE, RMSE, PICP, and MPIW represent the mean squared error, mean absolute error, interval coverage, and average interval width, respectively, and N is the number of prediction times included in the final prediction stage. In this embodiment, N = 48. t y' t Let It be the predicted value and the actual value of active power at time t, respectively, and let II(·) be the indicator function. t L t These are the upper and lower bounds of the power value distribution interval obtained from the probability distribution of active power, respectively; The obtained MAE, RMSE, PICP, and MPIW are compared with the first, second, third, and fourth thresholds, respectively. If the MAE does not exceed the first threshold, the RMSE does not exceed the second threshold, the PICP is not lower than the third threshold, and the MPIW does not exceed the fourth threshold, then the discrimination condition is considered to be met; otherwise, the discrimination condition is considered not met.

7. The wind farm power probability prediction method based on low-rank adaptation as described in claim 1, characterized in that, Step S4 specifically includes the following steps: S4.

1. Set the sampling stage and prediction stage with the current time as the reference time. Construct the feature vector of the sampling stage based on the unit power data and meteorological data of the past time. Construct the covariate vector of the prediction stage based on the meteorological forecast. Input the constructed feature vector and covariate vector into the time series prediction model to obtain the probability distribution of all prediction times. S4.

2. Establish a quantile prediction model. This model is used to determine the distribution interval of a value based on the probability distribution, and, combined with the set numerical distribution conditions, to obtain the position of a certain probability value within the distribution interval. This position is called the probability quantile. The quantile prediction model is described by the following formula: (3) Where P(·) is the probability function, q is the probability value, and y is the probability value. (q) t Let q be the probability quantile corresponding to the probability value q based on the probability distribution at time t. S4.

3. Set three probability values: lower limit, median, and upper limit, denoted as q. l q m q u , q l q m q u Substituting these values ​​into the quantile prediction model yields the corresponding probability quantile y. (ql) t y (qm) t y (qu) t , will y (ql) t and y (qu) t Using these as the upper and lower limits of the interval respectively, the quantile prediction interval [y] is obtained. (ql) t , y (qu) t ], y (qm) t As the predicted active power value at prediction time t, [y (ql) t , y (qu) t The active power prediction interval for prediction time t is used as the final output, which is the y value corresponding to all prediction times. (qm) t With [y (ql) t , y (qu) t ].

Images