Wind power point-interval prediction method and system based on improved RIME-ResBiLSTM-AM model

Abstract

The invention provides a wind power point-interval prediction intelligent adaptive regulation and control method and system based on an improved RIME-ResBiLSTM-AM model, and relates to the technical field of wind power prediction. The method comprises the following specific steps: in a plurality of acquisition periods, acquiring operation data and environment auxiliary data of a wind turbine generator, and generating standardized multi-dimensional time sequence data through preprocessing and period alignment; and training an improved RIME-ResBiLSTM-AM model by taking the preprocessed wind speed and wind direction time sequence data as input and historical wind power as output, and inputting predicted wind speed and wind direction data to obtain a wind power predicted value sequence. Extracting a power change rate and setting a dynamic threshold value, dividing the sequence into a stable section, a slow change section and a sudden change section, and respectively obtaining an error interval through adaptive bandwidth kernel density estimation, Bayesian quantile regression and an error cloud model; and superposing upper and lower limits of a corresponding error interval on the predicted value, and performing smooth connection through linear interpolation to obtain a wind power prediction interval.

Description

Technical Field

[0001] This invention relates to the field of wind power prediction technology, specifically to a point-interval prediction method and system for wind power based on an improved RIME-ResBiLSTM-AM model. Background Technology

[0002] Wind power, as a core renewable energy source, has achieved large-scale development. However, wind energy itself is highly random, intermittent, and volatile. Its large-scale grid connection poses severe challenges to the safe and stable operation of the power grid system, dispatch optimization, and renewable energy consumption. Accurate wind power forecasting has become a key technical support for solving this problem. Traditional point forecasting models have limited extraction capabilities for the coupled features of multiple variables such as wind speed and direction. In extreme weather or scenarios with drastic data fluctuations, the prediction error increases significantly. At the same time, existing interval forecasting methods lack differentiated modeling for different fluctuation stages of wind power, making it difficult to balance interval coverage and average interval width, and failing to quantify uncertainties under different scenarios. Therefore, developing a point-to-interval wind power forecasting technique based on an improved RIME-ResBiLSTM-AM model is particularly important.

[0003] BiLSTM models have become the mainstream infrastructure for wind power prediction due to their powerful ability to mine temporal dependencies. The integration of attention mechanisms further enhances the efficiency of key feature extraction. RIME algorithm, as an efficient multi-objective optimization algorithm, has the advantages of fast convergence speed and uniform solution distribution in parameter optimization, which can effectively solve the problem of model hyperparameter optimization. The residual connection structure of ResNet can alleviate the gradient vanishing problem in deep network training and improve the model's ability to fit complex features.

[0004] In the prior art, CN119362441A discloses a method and device for predicting offshore wind power points and intervals based on ICEEMDAN-VMD-LSTM-KDE. The specific steps are as follows: Two-layer decomposition to extract data features: First, the ICEEMDAN algorithm is used to decompose the original offshore wind speed time series to obtain multiple Intrinsic Mode Functions (IMFs), capturing dynamic changes in the data. For the high-frequency IMF1, which is difficult to predict after decomposition, the VMD algorithm is used for further decomposition to obtain more predictable sub-components. LSTM for point prediction: An LSTM neural network is used to predict all IMFs and sub-components obtained from the two decompositions, obtaining their respective sub-prediction results. All sub-prediction results are linearly summed to obtain the final offshore wind power point prediction value. KDE to determine the interval prediction range: The error between the point prediction result and the actual power value is calculated. KDE technology is used to estimate the probability density distribution of the prediction error. Combined with a given confidence level, the prediction interval (including upper and lower limits) of the offshore wind power is determined. However, this scheme has limited confidence level adaptation for interval prediction and does not consider the inaccuracy of the prediction interval caused by fluctuations between different segments of the point prediction.

[0005] The information disclosed in the background section is only intended to enhance the understanding of the background of this disclosure, and therefore may include information that does not constitute prior art known to those skilled in the art. Summary of the Invention

[0006] The purpose of this invention is to provide a wind power point-interval prediction method and system based on an improved RIME-ResBiLSTM-AM model to solve the problems mentioned in the background art.

[0007] To achieve the above objectives, the present invention provides the following technical solution: A wind power point-interval prediction method based on an improved RIME-ResBiLSTM-AM model, the specific steps of which include: Within multiple preset acquisition cycles, target wind turbine operation data and environmental auxiliary data are collected. The wind turbine operation data includes wind speed, wind direction and historical power. The environmental auxiliary data includes wind speed prediction and wind direction prediction for the next six hours. The collected data is preprocessed, and the preprocessed wind turbine operation data and environmental auxiliary data are aligned according to the acquisition cycle to obtain standardized multidimensional time series data. Furthermore, the preset multiple acquisition cycles are pre-defined continuous time intervals of equal length. The end time of each acquisition cycle is the same as the start time of the next acquisition cycle. All acquisition cycles are connected end to end on the time axis and are ordered continuously. The wind turbine operation data includes wind speed, wind direction and historical wind power. The environmental auxiliary data includes wind speed and wind direction prediction values ​​for the next 6 hours. Wind turbine operation data and environmental auxiliary data are collected continuously within each acquisition cycle. The collected data of each type are grouped according to their respective collection timestamps and time-aligned by collection period to form multi-dimensional time series data arranged by collection period. The multi-dimensional time series data includes wind speed time series data, wind direction time series data, historical power time series data, predicted wind speed time series data, and predicted wind direction time series data.

[0008] Furthermore, the preprocessing steps include first removing outliers from the collected data, and then denoising the data after outlier removal; the collected data is initially screened for reasonable value thresholds using the quartile method to remove outliers exceeding the thresholds; for the data after the initial outlier removal, the density-based spatial clustering and noise detection (DBSCAN) algorithm is used to identify anomalies through the density distribution of data points and remove hidden outliers in complex environments; and the data after outlier removal is denoised using the moving average smoothing method.

[0009] Furthermore, the logic for removing outliers is as follows: calculate the first quartile of the target data. and the third and fourth quartiles Calculate the interquartile range Set a lower limit for outlier detection thresholds. Upper limit Remove data values ​​that exceed the limit. For data points within a given range, if the upper limit of the data exceeds the actual physical upper limit, then the actual physical upper limit is used as the upper limit of the data. Select wind speed-power and wind direction-power data combinations to construct a data sample set, set the DBSCAN neighborhood radius, and the algorithm automatically divides densely connected samples into clusters and marks isolated points in low-density areas as outliers. The denoising logic is as follows: set the window length, use a single-sided window for samples at the window edge, and calculate the mean of all samples in the window before and after each data point as the smoothed value.

[0010] The wind speed and wind direction time series data from the preprocessed wind turbine operation data are used as inputs to the improved RIME-ResBiLSTM-AM model, and the historical wind power time series data are used as outputs to train the improved RIME-ResBiLSTM-AM model. The predicted wind speed and wind direction time series data are then input into the trained improved RIME-ResBiLSTM-AM model to obtain the wind power prediction sequence.

[0011] Furthermore, wind speed and wind direction time-series data are used as inputs to the improved RIME-ResBiLSTM-AM model, and historical wind power time-series data are used as outputs. The mean square error between the output power and the historical power values ​​is calculated as the loss function. ,in Historical power For the output power, the loss function is fed back into the improved model until... Training is terminated.

[0012] The power change rate is extracted from the wind power prediction value sequence by using a sliding window combined with the first-order difference method. A dynamic threshold is set to identify the turning point of power fluctuation, and the wind power prediction value sequence is divided into a stable segment, a slowly changing segment, and a sudden change segment. Furthermore, the method for dividing the wind power prediction sequence into stationary, gradually changing, and abruptly changing segments is as follows: a sliding window is used to process the predicted power sequence, and the first-order difference within the window, i.e., the power change rate, is calculated. ,in for Power at any moment for For the power at any given time, kernel density estimation is used to fit the distribution of the power change rate, and the 25th percentile is taken as the low threshold. The 75th percentile was used as the high threshold. The predicted power prediction sequence is segmented based on a dynamic threshold, with the stationary segment being: Gradual transition phase: Mutation segment: Record the time points for each segment, and calculate the error between the predicted and measured values ​​for each fluctuation segment based on historical data. .

[0013] For the stationary power sequence, an adaptive bandwidth kernel density estimation method is used to fit the error distribution, accurately obtaining the error interval of the stationary segment. An error model based on Bayesian quantile regression is constructed. For the gradually varying power sequence, the prior distribution of the error and sample likelihood information are integrated through Bayes' theorem to estimate multiple quantiles, obtaining a reliable error boundary for the gradually varying power sequence. The power sequence of the abrupt change segment is input into the error cloud model, and the error interval of the power sequence of the abrupt change segment is output. Furthermore, the method for interval estimation of the stationary segment is as follows: the kernel function for adaptive bandwidth kernel density estimation is selected as a Gaussian kernel. Adaptive bandwidth calculation: Set the bandwidth for each error sample. ,in The standard deviation of the local error. The overall error standard deviation; error distribution fitting: by obtaining the error probability density function. Solve for the condition based on a 95% confidence level. of and This serves as the upper and lower limits of the error range for the steady-state segment, forming the error range of the steady-state segment. The method for interval estimation of the gradually changing segment is as follows: select the 0.05 and 0.95 quantiles to construct a quantile regression model. ,in The values ​​are 0.05 and 0.95. For wind speed change rate, For wind direction change rate, The power change rate is used to solve for the model parameters by minimizing the weighted absolute error loss function. Substituting the wind speed change rate, wind direction change rate, and power change rate corresponding to the gradual change phase into the model, we obtain the lower limit of error corresponding to the 0.05 quantile. The upper limit of error corresponding to the 0.95 quantile Forming the error range of the gradual change segment The method for interval estimation of the mutation segment is as follows: K-means is used to evaluate the error sample set of the mutation segment. Clustering, with the K-value determined using the elbow rule, divides the errors into three error clusters with similar distribution characteristics: a small-amplitude random error cluster, a medium-amplitude random error cluster, and an extreme error cluster. For each error cluster, three core parameters of the cloud model are calculated: the mean error within the cluster. Intra-cluster error standard deviation Standard deviation of intra-cluster error Construct cloud models corresponding to each cluster; for abrupt change segments, use cosine similarity to calculate the similarity between their power fluctuation amplitude and each error cluster, and select the similarity... Attached Figure Description

[0014] Figure 1 This is a schematic diagram of the overall method flow of the present invention; Figure 2 A graph showing the actual power versus the upper and lower limits of the predicted range; Figure 3 This is a structural block diagram of the overall system of the present invention. Detailed Implementation

[0015] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to specific embodiments.

[0016] It should be noted that, unless otherwise defined, the technical or scientific terms used in this invention should have the ordinary meaning understood by one of ordinary skill in the art to which this invention pertains. The terms "first," "second," and similar terms used in this invention do not indicate any order, quantity, or importance, but are merely used to distinguish different components. Terms such as "comprising" or "including" mean that the element or object preceding the word encompasses the elements or objects listed following the word and their equivalents, without excluding other elements or objects. Terms such as "connected" or "linked" are not limited to physical or mechanical connections, but can include electrical connections, whether direct or indirect. Terms such as "upper," "lower," "left," and "right" are used only to indicate relative positional relationships; when the absolute position of the described object changes, the relative positional relationship may also change accordingly.

[0017] Example: Please see Figure 1 The present invention provides a technical solution: A wind power point-interval prediction method based on an improved RIME-ResBiLSTM-AM model, the specific steps of which include: Step 1: Collect target wind turbine operation data and environmental auxiliary data within multiple preset collection periods. The wind turbine operation data includes wind speed, wind direction and historical power. The environmental auxiliary data includes wind speed prediction and wind direction prediction for the next six hours. Preprocess the collected data and align the preprocessed wind turbine operation data and environmental auxiliary data according to the collection period to obtain standardized multidimensional time series data. The preset multiple acquisition cycles are pre-defined continuous time intervals of equal length. The end time of each acquisition cycle is the same as the start time of the next acquisition cycle. All acquisition cycles are connected end to end on the time axis and are ordered continuously. The wind turbine operation data includes wind speed, wind direction and historical wind power. The environmental auxiliary data includes wind speed and wind direction prediction values ​​for the next 6 hours. Wind turbine operation data and environmental auxiliary data are collected continuously within each acquisition cycle. The collected data of each type are grouped according to their respective collection timestamps and time-aligned by collection period to form multi-dimensional time series data arranged by collection period. The multi-dimensional time series data includes wind speed time series data, wind direction time series data, historical power time series data, predicted wind speed time series data, and predicted wind direction time series data.

[0018] The preprocessing steps include first removing outliers from the collected data, and then denoising the data after outlier removal. The collected data is initially screened for reasonable value thresholds using the quartile method to remove outliers exceeding the thresholds. For the data after initial outlier removal, the density-based spatial clustering and noise detection (DBSCAN) algorithm is used to identify anomalies through the density distribution of data points, removing hidden outliers in complex environments. Finally, the data after outlier removal is denoised using a moving average smoothing method.

[0019] The logic for removing outliers is as follows: calculate the first quartile of the target data. and the third and fourth quartiles Calculate the interquartile range Set a lower limit for outlier detection thresholds. Upper limit Remove data values ​​that exceed the limit. For data points within a given range, if the upper limit of the data exceeds the actual physical upper limit, then the actual physical upper limit is used as the upper limit of the data. Select wind speed-power and wind direction-power data combinations to construct a data sample set, set the DBSCAN neighborhood radius, and the algorithm automatically divides densely connected samples into clusters and marks isolated points in low-density areas as outliers. The denoising logic is as follows: set the window length, use a single-sided window for samples at the window edge, and calculate the mean of all samples in the window before and after each data point as the smoothed value.

[0020] The collected data encompasses wind power operation data and environmental auxiliary data, including core parameters reflecting the historical operating status of the units as well as environmental prediction information affecting future output. This forms a dual-dimensional data system, addressing the limitations of traditional forecasting that relies solely on historical data and ignores short-term environmental trends. A dual screening mechanism using quartiles and the DBSCN algorithm is employed to accurately identify different types of outliers. After outlier removal, a moving average smoothing method is used. By setting the window length and processing edge samples with single-sided windows, the data more closely reflects actual wind speed and power variation trends, while avoiding the loss or distortion of edge data due to incomplete windows. This ensures the continuity and integrity of the entire time-series dataset, supporting subsequent model processing of full-time data. Through multi-dimensional data acquisition, precise preprocessing, and time alignment, data quality is guaranteed from the source, providing reliable input for subsequent model training and prediction.

[0021] Step 2: Use the preprocessed wind speed and wind direction time series data from the wind turbine operation data as input to the improved RIME-ResBiLSTM-AM model, and use the historical wind power time series data as the model output to train the improved RIME-ResBiLSTM-AM model; input the predicted wind speed and predicted wind direction time series data into the trained improved RIME-ResBiLSTM-AM model to obtain the wind power prediction sequence; use the wind speed and wind direction time series data as input to the improved RIME-ResBiLSTM-AM model, and use the historical wind power time series data as the model output, calculate the mean square error between the output power and the historical power value as the loss function. ,in Historical power For the output power, the loss function is fed back into the improved model until... Training terminates. The predicted wind speed and predicted wind direction time-series data are then input into the trained improved RIME-ResBiLSTM-AM model to obtain the wind power prediction sequence.

[0022] Using wind speed and wind direction time-series data as inputs and historical wind power time-series data as outputs, the training objective is designed directly targeting the core influencing factors of wind power, improving model learning efficiency. The improved RIME-ResBiLSTM-AM model integrates the temporal feature capture capability of ResBiLSTM with the key feature focusing capability of AM, while also incorporating optimized features from the RIME algorithm. This structure effectively solves the problems of gradient vanishing and insufficient weights of key features in traditional LSTM models with long-term data. ResBiLSTM ensures feature transfer from long-term data through residual connections, and the bidirectional structure utilizes both past and future temporal information. AM automatically increases the weights of features with the greatest impact on power, improving the accuracy of the prediction model. Using mean squared error (MSE) as the loss function amplifies the penalty for larger errors, avoiding distortion in error assessment caused by differences in power value ranges. This allows model training to focus more on reducing the deviation of actual power prediction. With MSE < 3% as the training termination criterion, model training is strictly controlled from a quantitative perspective to ensure that the model has achieved a high fitting accuracy during the training phase, avoiding overfitting or underfitting. This provides a guarantee for outputting low-error power point prediction values ​​in the subsequent prediction phase, outputting a sequence of point prediction values ​​with a deviation of less than 3% from the actual power, thus meeting the basic requirements of wind power dispatch for power prediction accuracy.

[0023] Step 3: Extract the power change rate from the wind power prediction value sequence using a sliding window combined with the first-order difference method, set a dynamic threshold to identify the turning point of power fluctuation, and divide the wind power prediction value sequence into a stable segment, a slowly changing segment, and a sudden change segment. The method for dividing the wind power prediction sequence into stationary, gradually changing, and abruptly changing segments is as follows: A sliding window is used to process the predicted power sequence, and the first-order difference within the window, i.e., the power change rate, is calculated. ,in for Power at any moment for For the power at any given time, kernel density estimation is used to fit the distribution of the power change rate, and the 25th percentile is taken as the low threshold. The 75th percentile was used as the high threshold. The predicted power prediction sequence is segmented based on a dynamic threshold, with the stationary segment being: Gradual transition phase: Mutation segment: Record the time points for each segment, and calculate the error between the predicted and measured values ​​for each fluctuation segment based on historical data. .

[0024] A sliding window approach is used to process the predicted power sequence, calculating the power change rate within the window. The sliding window avoids misjudgments caused by data fluctuations at a single moment. First-order differencing transforms absolute power changes into relative change rates, more intuitively quantifying the intensity of power fluctuations and providing a quantifiable basis for subsequent segmentation. Kernel density estimation is used to fit the distribution of the power change rate, employing a dynamic threshold instead of a fixed one. This dynamic threshold adapts to the power fluctuation characteristics of different wind fields and time periods, ensuring a match between the threshold and actual fluctuations. Based on the dynamic threshold, the power sequence is divided into stationary, gradually changing, and abruptly changing segments. This segmentation allows for targeted interval estimation, selecting appropriate methods based on the error characteristics of different fluctuation segments, significantly improving the accuracy and reliability of interval predictions. Simultaneously, time nodes for each segment are recorded, and the error between the predicted and measured values ​​is calculated, providing an error sample basis for subsequent interval estimation.

[0025] Step 4: For the stationary power sequence, use adaptive bandwidth kernel density estimation to fit the error distribution and accurately obtain the error interval of the stationary segment; construct an error model based on Bayesian quantile regression, and use Bayes' theorem to fuse the prior distribution of error and sample likelihood information for the gradually varying power sequence to estimate multiple quantiles and obtain a reliable error boundary for the gradually varying power sequence; input the power sequence of the abrupt change segment into the error cloud model and output the error interval of the power sequence of the abrupt change segment. The method for estimating the interval of the stationary segment is as follows: the kernel function for adaptive bandwidth kernel density estimation is a Gaussian kernel. Adaptive bandwidth calculation: Set the bandwidth for each error sample. ,in The standard deviation of the local error. The overall error standard deviation; error distribution fitting: by obtaining the error probability density function. Solve for the condition based on a 95% confidence level. of and This serves as the upper and lower limits of the error range for the steady-state segment, forming the error range of the steady-state segment. In the stationary phase, power fluctuations are small, and the error distribution is relatively concentrated. Accurate fitting of the error distribution is achieved through adaptive bandwidth and a Gaussian kernel. The Gaussian kernel's characteristic distribution properties highly match the stationary phase error distribution, which is concentrated near the mean and decreases at both ends. This avoids distribution distortion caused by non-smooth kernel functions, ensuring the reasonableness of the error distribution fit. Traditional kernel density estimation uses a fixed bandwidth, which easily leads to over-smoothing in dense regions and under-smoothing in sparse regions. However, by allocating an adaptive bandwidth to each error sample, spurious peaks are avoided, and the resulting error probability density function more accurately reflects the true distribution of errors in the stationary phase.

[0026] The method for interval estimation of the gradually changing segment is as follows: select the 0.05 and 0.95 quantiles to construct a quantile regression model. ,in The values ​​are 0.05 and 0.95. For wind speed change rate, For wind direction change rate, The power change rate is used to solve for the model parameters by minimizing the weighted absolute error loss function. Substituting the wind speed change rate, wind direction change rate, and power change rate corresponding to the gradual change phase into the model, we obtain the lower limit of error corresponding to the 0.05 quantile. The upper limit of error corresponding to the 0.95 quantile Forming the error range of the gradual change segment Quantile regression models directly output the dynamic factors affecting error, capturing the pattern that the greater the rate of change, the greater the potential error. This allows the error interval to dynamically adjust with the rate of change, avoiding the problem of traditional fixed intervals failing to adapt to dynamic errors. Bayesian fusion, by minimizing the weighted absolute error loss function to solve for model parameters, fuses the prior distribution of the error. Compared to traditional quantile regression, which relies solely on sample data, the Bayesian method utilizes prior information to compensate for insufficient sample size, reducing the uncertainty of parameter estimation. This makes the lower and upper bounds of the error more reliable, ultimately resulting in a gradually changing error interval that covers 95% of the actual error and accurately reflects the dynamic trend of error with the rate of change.

[0027] The method for interval estimation of the mutation segment is as follows: K-means is used to evaluate the error sample set of the mutation segment. Clustering, with the K-value determined using the elbow rule, divides the errors into three error clusters with similar distribution characteristics: a small-amplitude random error cluster, a medium-amplitude random error cluster, and an extreme error cluster. For each error cluster, three core parameters of the cloud model are calculated: the mean error within the cluster. Intra-cluster error standard deviation Standard deviation of intra-cluster error The cloud model corresponding to each cluster is constructed. For abrupt changes, the cosine similarity is used to calculate the similarity between the power fluctuation amplitude and each error cluster. The cloud model corresponding to the error cluster with the highest similarity is selected to generate an error interval with a 95% confidence level. K-means clustering can separate extreme errors from regular errors, avoiding interference from extreme errors in fitting the overall error distribution and solving the problem of excessively wide intervals or insufficient coverage caused by extreme values ​​in traditional methods. For each error cluster, the core parameters of the cloud model are calculated, and a cloud model corresponding to each cluster is constructed. The cloud model can simultaneously describe the central tendency, dispersion, and uncertainty of the error, and is particularly suitable for the characteristics of strong randomness and high uncertainty in abrupt change segments. Compared with traditional probability distributions, it can more accurately characterize the distribution characteristics of extreme errors. For the current abrupt change segment, the similarity between its power fluctuation amplitude and each error cluster is calculated using cosine similarity. The cloud model corresponding to the cluster with the highest similarity is selected to generate an error interval with a 95% confidence level. This dynamic matching ensures that the interval estimation can adapt to different types of abrupt change scenarios. Even if extreme errors occur, the interval can effectively cover the data, ensuring the ability of the interval to resist interference from extreme situations in the abrupt change segment.

[0028] Step 5: Superimpose the wind power prediction values ​​with the upper and lower limits of the error interval of the corresponding fluctuation segment to obtain the preliminary prediction interval of each segment. Use linear interpolation to smooth the connection between the segment intervals to obtain the wind power prediction interval.

[0029] The method for synthesizing the error intervals of the stable segment, the gradually changing segment, and the abrupt change segment to obtain the prediction interval is as follows: The error intervals corresponding to the predicted values ​​of each segment are superimposed to obtain the preliminary prediction interval: [Stable segment prediction interval] Prediction interval of gradual change segment Mutation segment prediction interval Linear interpolation is used to smooth the upper and lower limits of the intervals at the junctions to obtain the wind power prediction intervals. Interval overlay directly combines the low-error point predictions with the error intervals of each adapted segment. This retains the accuracy advantage of point predictions while quantifying the uncertainty of predictions through error intervals, upgrading the results from single numerical values ​​to numerical ranges and confidence levels, which meets the actual needs of wind power dispatch. Since the error interval widths and upper and lower limits of each segment may differ, direct splicing will cause discontinuities at the interval junctions, affecting the practical application value of the intervals. Using linear interpolation to smooth the upper and lower limits of the intervals at the junctions can eliminate the junction discontinuities and improve the usability of the results.

[0030] As shown in Table 1, the data in Table 1 reflects the wind power prediction data within a portion of the data collection period.

[0031] Table 1 shows the point prediction error. The mean squared error (MSE) meets the termination threshold set by the scheme. At a 95% confidence level, the actual power falls within the upper and lower limits of the interval prediction, the interval coverage reaches 100%, and after linear interpolation at the segment connection, there are no interval breaks or abrupt changes, and the data continuity is strong.

[0032] Please see Figure 3 The present invention also provides a wind power point-interval prediction system based on an improved RIME-ResBiLSTM-AM model. This improved RIME-ResBiLSTM-AM model-based wind power point-interval prediction system is used to implement the wind power point-interval prediction method described above, and includes: Data acquisition module: used to collect target wind turbine operating data and environmental auxiliary data within multiple preset acquisition cycles. The wind turbine operating data includes wind speed, wind direction and historical power. The environmental auxiliary data includes wind speed forecast and wind direction forecast for the next six hours. Data preprocessing module: Used to preprocess the collected data, aligning the preprocessed wind turbine operation data and environmental auxiliary data according to the collection period to obtain standardized multidimensional time-series data; Model training module: Used to take the wind speed time series data and wind direction time series data in the preprocessed wind turbine operation data as input terms of the improved RIME-ResBiLSTM-AM model, and take the historical wind power time series data as the output terms of the model to train the improved RIME-ResBiLSTM-AM model; Point prediction calculation module: used to input the predicted wind speed time series data and the predicted wind direction time series data into the trained improved RIME-ResBiLSTM-AM model to obtain the wind power prediction value sequence; Power prediction sequence segmentation module: It is used to extract the power change rate of the wind power prediction value sequence by using a sliding window combined with the first-order difference method, set a dynamic threshold to identify the turning point of power fluctuation, and divide the wind power prediction value sequence into a stable segment, a slowly changing segment, and a sudden change segment. The interval prediction calculation module is used to: fit the error distribution of the power sequence in the stationary segment using adaptive bandwidth kernel density estimation, and accurately obtain the error interval of the stationary segment; construct an error model based on Bayesian quantile regression, and estimate multiple quantiles for the power sequence in the gradually changing segment by fusing the prior distribution of error and sample likelihood information through Bayes' theorem, thereby obtaining a reliable error boundary for the power sequence in the gradually changing segment; input the power sequence in the abrupt segment into the error cloud model, and output the error interval of the power sequence in the abrupt segment; superimpose the wind power prediction value with the upper and lower limits of the error interval of the corresponding fluctuation segment to obtain the preliminary prediction interval for each segment, and use linear interpolation to smooth the connection between the segment intervals to obtain the wind power prediction interval. The above formulas are all dimensionless and numerical calculations. The formulas are derived from software simulation based on a large amount of collected data to obtain the most recent real situation. The preset parameters in the formulas are set by those skilled in the art according to the actual situation.

[0033] The above embodiments can be implemented, in whole or in part, by software, hardware, firmware, or any other combination thereof. When implemented in software, the above embodiments can be implemented, in whole or in part, as a computer program product. Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented by electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution.

[0034] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0035] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application.

Claims

1. A wind power point-interval prediction method based on an improved RIME-ResBiLSTM-AM model, characterized in that, Includes the following steps: Within multiple preset acquisition cycles, target wind turbine operation data and environmental auxiliary data are collected. The wind turbine operation data includes wind speed, wind direction and historical power. The environmental auxiliary data includes wind speed prediction and wind direction prediction for the next six hours. The collected data is preprocessed, and the preprocessed wind turbine operation data and environmental auxiliary data are aligned according to the acquisition cycle to obtain standardized multidimensional time series data. The wind speed and wind direction time series data from the preprocessed wind turbine operation data are used as inputs to the improved RIME-ResBiLSTM-AM model, and the historical wind power time series data are used as outputs to train the improved RIME-ResBiLSTM-AM model. The predicted wind speed and predicted wind direction time series data are then input into the trained improved RIME-ResBiLSTM-AM model to obtain the wind power prediction sequence. The power change rate is extracted from the wind power prediction value sequence by using a sliding window combined with the first-order difference method. A dynamic threshold is set to identify the turning point of power fluctuation, and the wind power prediction value sequence is divided into a stable segment, a slowly changing segment, and a sudden change segment. For the stationary power sequence, an adaptive bandwidth kernel density estimation method is used to fit the error distribution, accurately obtaining the error interval of the stationary segment. An error model based on Bayesian quantile regression is constructed. For the gradually varying power sequence, the prior distribution of the error and sample likelihood information are integrated through Bayes' theorem to estimate multiple quantiles, obtaining a reliable error boundary for the gradually varying power sequence. The power sequence of the abrupt change segment is input into the error cloud model, and the error interval of the power sequence of the abrupt change segment is output. The wind power prediction values ​​are superimposed with the upper and lower limits of the error range of the corresponding fluctuation segment to obtain the preliminary prediction range of each segment. The connection between the segment intervals is smoothed by linear interpolation to obtain the wind power prediction range.

2. The wind power point-interval prediction method based on the improved RIME-ResBiLSTM-AM model according to claim 1, characterized in that: The preset multiple acquisition cycles are pre-defined continuous time intervals of equal length. The end time of each acquisition cycle is the same as the start time of the next acquisition cycle. All acquisition cycles are connected end to end on the time axis and are ordered continuously. The wind turbine operation data includes wind speed, wind direction and historical wind power. The environmental auxiliary data includes wind speed and wind direction prediction values ​​for the next 6 hours. Wind turbine operation data and environmental auxiliary data are collected continuously within each acquisition cycle. The collected data of each type are grouped according to their respective collection timestamps and time-aligned by collection period to form multi-dimensional time series data arranged by collection period. The multi-dimensional time series data includes wind speed time series data, wind direction time series data, historical power time series data, predicted wind speed time series data, and predicted wind direction time series data.

3. The wind power point-interval prediction method based on the improved RIME-ResBiLSTM-AM model according to claim 1, characterized in that: The preprocessing steps include first removing outliers from the collected data, and then denoising the data after outlier removal. The collected data is initially screened for reasonable value thresholds using the quartile method to remove outliers exceeding the thresholds. For the data after initial outlier removal, the density-based spatial clustering and noise detection (DBSCAN) algorithm is used to identify anomalies through the density distribution of data points, removing hidden outliers in complex environments. Finally, the data after outlier removal is denoised using a moving average smoothing method.

4. The wind power point-interval prediction method based on the improved RIME-ResBiLSTM-AM model according to claim 3, characterized in that: The logic for removing outliers is as follows: calculate the first quartile of the target data. and the third and fourth quartiles Calculate the interquartile range Set a lower limit for outlier detection thresholds. Upper limit Remove data values ​​that exceed the limit. For data points within a range, if the upper limit of the data exceeds the actual physical upper limit, then the actual physical upper limit shall be used as the upper limit of the data. Select wind speed-power and wind direction-power data combinations to construct a data sample set. Set the DBSCAN neighborhood radius. The algorithm automatically divides densely connected samples into clusters and marks isolated points in low-density areas as outliers. The denoising logic is as follows: set the window length, use a single-sided window for samples at the window edge, and calculate the mean of all samples in the window before and after each data point as the smoothed value.

5. The wind power point-interval prediction method based on the improved RIME-ResBiLSTM-AM model according to claim 1, characterized in that: Wind speed and wind direction time-series data are used as inputs to the improved RIME-ResBiLSTM-AM model, and historical wind power time-series data are used as outputs. The mean square error between the output power and the historical power values ​​is calculated as the loss function. ,in Historical power For the output power, the loss function is fed back into the improved model until... Training terminates; the predicted wind speed time series data and predicted wind direction time series data are input into the trained improved RIME-ResBiLSTM-AM model to obtain the wind power prediction value sequence.

6. The wind power point-interval prediction method based on the improved RIME-ResBiLSTM-AM model according to claim 1, characterized in that: The method for dividing the wind power prediction sequence into stationary, gradually changing, and abruptly changing segments is as follows: A sliding window is used to process the predicted power sequence, and the first-order difference within the window, i.e., the power change rate, is calculated. ,in for Power at any moment for For the power at any given time, kernel density estimation is used to fit the distribution of the power change rate, and the 25th percentile is taken as the low threshold. The 75th percentile was used as the high threshold. The predicted power prediction sequence is segmented based on a dynamic threshold, with the stationary segment being: Gradual transition phase: Mutation segment: Record the time points for each segment, and calculate the error between the predicted and measured values ​​for each fluctuation segment based on historical data. .

7. The wind power point-interval prediction method based on the improved RIME-ResBiLSTM-AM model according to claim 1, characterized in that: The method for estimating the interval of the stationary segment is as follows: the kernel function for adaptive bandwidth kernel density estimation is a Gaussian kernel. ; Adaptive bandwidth calculation: Set the bandwidth for each error sample. ,in The standard deviation of the local error. The overall error standard deviation; error distribution fitting: by obtaining the error probability density function. Solve for the condition based on a 95% confidence level. of and This serves as the upper and lower limits of the error range for the steady-state segment, forming the error range of the steady-state segment. ; The method for interval estimation of the gradually changing segment is as follows: select the 0.05 and 0.95 quantiles to construct a quantile regression model. ,in The values ​​are 0.05 and 0.

95. For wind speed change rate, For wind direction change rate, The power change rate is used to solve for the model parameters by minimizing the weighted absolute error loss function. Substituting the wind speed change rate, wind direction change rate, and power change rate corresponding to the gradual change phase into the model, we obtain the lower limit of error corresponding to the 0.05 quantile. The upper limit of error corresponding to the 0.95 quantile Forming the error range of the gradual change segment ; The method for interval estimation of the mutation segment is as follows: K-means is used to evaluate the error sample set of the mutation segment. Clustering, with the K-value determined using the elbow rule, divides the errors into three error clusters with similar distribution characteristics: a small-amplitude random error cluster, a medium-amplitude random error cluster, and an extreme error cluster. For each error cluster, three core parameters of the cloud model are calculated: the mean error within the cluster. Intra-cluster error standard deviation Standard deviation of intra-cluster error The cloud model corresponding to each cluster is constructed. For abrupt changes, the cosine similarity is used to calculate the similarity between the power fluctuation amplitude and each error cluster. The cloud model corresponding to the error cluster with the highest similarity is selected to generate an error interval with a 95% confidence level. .

8. A wind power point-interval prediction method based on an improved RIME-ResBiLSTM-AM model according to claim 7, characterized in that... The method for synthesizing the error intervals of the stable segment, the gradually changing segment, and the abruptly changing segment to obtain the prediction interval is as follows: The error intervals corresponding to the predicted values ​​of each segment are superimposed to obtain the preliminary prediction interval: Stable segment prediction interval. Prediction interval of gradual change segment Mutation segment prediction interval The wind power prediction range is obtained by smoothing the upper and lower limits of the interval at the junction using linear interpolation.

9. A wind power point-interval prediction system based on an improved RIME-ResBiLSTM-AM model, the system being used to implement the wind power point-interval prediction method based on the improved RIME-ResBiLSTM-AM model as described in any one of claims 1-8, characterized in that: Data acquisition module: used to collect target wind turbine operating data and environmental auxiliary data within multiple preset acquisition cycles. The wind turbine operating data includes wind speed, wind direction and historical power. The environmental auxiliary data includes wind speed forecast and wind direction forecast for the next six hours. Data preprocessing module: Used to preprocess the collected data, aligning the preprocessed wind turbine operation data and environmental auxiliary data according to the collection period to obtain standardized multidimensional time-series data; Model training module: Used to take the wind speed time series data and wind direction time series data in the preprocessed wind turbine operation data as input terms of the improved RIME-ResBiLSTM-AM model, and take the historical wind power time series data as the output terms of the model to train the improved RIME-ResBiLSTM-AM model; Point prediction calculation module: used to input the predicted wind speed time series data and the predicted wind direction time series data into the trained improved RIME-ResBiLSTM-AM model to obtain the wind power prediction value sequence; Power prediction sequence segmentation module: It is used to extract the power change rate of the wind power prediction value sequence by using a sliding window combined with the first-order difference method, set a dynamic threshold to identify the turning point of power fluctuation, and divide the wind power prediction value sequence into a stable segment, a slowly changing segment, and a sudden change segment. The interval prediction calculation module is used to: fit the error distribution of the power sequence in the stationary segment using adaptive bandwidth kernel density estimation, and accurately obtain the error interval of the stationary segment; construct an error model based on Bayesian quantile regression, and estimate multiple quantiles for the power sequence in the gradually changing segment by fusing the prior distribution of the error with sample likelihood information using Bayes' theorem, and obtain a reliable error boundary for the power sequence in the gradually changing segment; input the power sequence in the abrupt segment into the error cloud model, and output the error interval of the power sequence in the abrupt segment; superimpose the wind power prediction values ​​with the upper and lower limits of the error interval of the corresponding fluctuation segment to obtain the preliminary prediction interval for each segment, and use linear interpolation to smooth the connection between the segment intervals to obtain the wind power prediction interval.

Images