Ultra-short-term wind power prediction method and system based on aurora optimization algorithm and error correction

Abstract

The invention relates to an ultra-short-term wind power prediction method and system based on an aurora optimization algorithm and error correction, and the method comprises the steps: obtaining historical power data of a target wind power station before a target prediction time period and meteorological feature data of the same period, and carrying out the normalization processing, and forming an input feature vector; constructing a CNN-BiGRU-Attention model which takes the input feature vector as the input and takes the preliminarily predicted wind power as the output; performing adaptive optimization on hyper-parameters of the model based on an aurora optimization algorithm; calculating a prediction error sequence and performing pattern recognition based on the preliminarily predicted wind power output by the optimized model and the pre-acquired real power to obtain a pattern type; calling a corresponding pre-constructed and trained error correction sub-model, and taking the input feature vector and the error sequence as input and output power correction values; and obtaining a final ultra-short-term wind power prediction value based on the preliminarily predicted wind power and the power correction value.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of wind power prediction, in particular to a method and system for ultra-short-term wind power prediction based on aurora optimization algorithm and error correction. BACKGROUND

[0002] The installed capacity of wind power in power systems continues to grow, however, wind power output is affected by low-altitude wind speed, wind direction, air pressure, turbulence intensity and other meteorological factors, and has obvious randomness, volatility and intermittency. Under complex weather conditions, such as local convection, frontal passage, low-level jet, etc., wind speed will change rapidly on a minute time scale, causing severe fluctuations in wind power. Such high-frequency random fluctuations pose a great challenge to real-time dispatching and stability control of the power grid. Therefore, achieving minute-level ultra-short-term wind power prediction has become an important technical problem to ensure the safe operation of the power grid and improve the ability of new energy consumption.

[0003] Existing wind power prediction methods mainly include physical modeling methods and data-driven methods. Physical modeling methods rely on numerical weather prediction (NWP) to simulate future wind fields through atmospheric dynamics equations, and can predict power changes on an hourly or longer time scale. However, in minute-level high-frequency prediction, the prediction lag problem is prominent due to limitations in initial field accuracy and grid resolution. Data-driven methods use historical wind speed, wind direction, power and other data to establish a power prediction model through statistical modeling or machine learning methods, which improves prediction accuracy to some extent. However, such methods perform poorly when faced with high-frequency fluctuations, sudden wind speed changes and complex nonlinear relationships, especially in minute-level prediction scenarios, where there is still a large prediction error.

[0004] In recent years, with the development of deep learning technology and intelligent optimization algorithms, high-precision ultra-short-term wind power prediction methods based on multi-source meteorological feature fusion have gradually become a research hotspot. Although certain progress has been made in terms of accuracy and stability, in the minute-level ultra-short-term prediction scenario, existing methods still have performance bottlenecks when faced with severe wind speed fluctuations and complex wind field characteristics, and there is an urgent need for more accurate and robust prediction methods.

[0005] The main methods used in existing technologies are: 1) Calculate future wind speed, wind direction and other meteorological parameters using a high-resolution numerical weather prediction model, and then estimate future power output based on the wind turbine power curve. This method can achieve certain results in medium and long-term prediction, but it lags behind in responding to local meteorological changes and minute-level wind speed mutations, and relies on the accuracy of meteorological initial fields and boundary conditions, resulting in large prediction errors under complex weather conditions.

[0006] 2) Using historical wind speed, wind direction and power data, the prediction is made by constructing a statistical regression or machine learning model, which improves the ultra-short-term prediction accuracy to a certain extent. However, this method has limited ability to extract high-dimensional meteorological features when dealing with non-stationary wind fields and sudden power fluctuations, and it is difficult to fully utilize multi-source information, especially in low-altitude wind shear, local vortex and other special meteorological conditions.

[0007] 3) Local key features are extracted using convolutional neural network (CNN), time-dependent relationships are captured using gated recurrent neural network, and attention mechanism is introduced to dynamically allocate feature weights. This method has great advantages in high-dimensional feature fusion and non-linear relationship modeling, but still has two shortcomings: one is the high sensitivity of model hyperparameters, which is difficult to adapt to complex scenarios of different wind farms; the second is that there may still be systematic bias in the preliminary prediction results, and the adaptability to minute-level dramatic fluctuations is insufficient.

[0008] In summary, the wind power prediction based on multi-source data fusion has become a research hotspot at home and abroad in recent years, but there are still challenges in complex wind field feature extraction, minute-level mutation prediction mechanism establishment, and prediction bias adaptive correction, so it is urgent to propose a high-precision and high-robustness ultra-short-term wind power prediction method. SUMMARY

[0009] The purpose of the present application is to overcome the defects of the prior art and provide an ultra-short-term wind power prediction method and system based on aurora optimization algorithm and error correction, to solve or partially solve the problems of non-linear modeling between feature and wind power, weak generalization ability of model hyperparameter tuning, lack of error correction mechanism and other problems.

[0010] The purpose of the present application can be realized by the following technical solutions: In one aspect of the present application, an ultra-short-term wind power prediction method based on aurora optimization algorithm and error correction is provided, comprising the following steps: Obtain historical power data of the target wind farm station before the target prediction period and the synchronous meteorological feature data and perform normalization processing to form an input feature vector; Construct a CNN-BiGRU-Attention model with the input feature vector as input and the preliminary predicted wind power as output; Adaptively optimize the hyperparameters of the CNN-BiGRU-Attention model based on the aurora optimization algorithm; Based on the preliminary predicted wind power output by the optimized CNN-BiGRU-Attention model and the pre-acquired real power, calculate the prediction error sequence and perform pattern recognition to obtain the pattern type; Based on the aforementioned mode type, the corresponding pre-built and trained error correction sub-model is invoked, taking the input feature vector and error sequence as input, and outputting the power correction value; Based on the preliminary wind power forecast and the power correction value, the final ultra-short-term wind power forecast value is obtained.

[0011] As a preferred technical solution, the process of obtaining the input feature vector includes the following steps: Historical power data of the target wind farm before the target prediction period and meteorological characteristic data of the same period are obtained, wherein the meteorological characteristics include irradiance, temperature, humidity, wind speed, wind direction and cloud cover; Historical power data and meteorological feature data are time-aligned and missing values ​​are processed to form an input feature vector; The input feature vector is normalized.

[0012] As a preferred technical solution, the CNN-BiGRU-Attention model is modeled as follows: The input feature vector is tensor-encoded and fed into the built-in convolutional neural network, which extracts local spatiotemporal key features through multi-layer convolution and pooling operations. The high-dimensional feature maps output by the built-in convolutional neural network are input into the built-in bidirectional gated recurrent unit network to capture temporal dependencies from both forward and backward time dimensions and extract bidirectional dynamic features of the input sequence. Attention processing is performed on the hidden state sequence output by the bidirectional gated recurrent unit network, and weights are assigned to the importance of the predicted target according to the features at each time step to construct a weighted feature representation; The preliminary predicted wind power output is based on the weighted feature representation.

[0013] As a preferred technical solution, the adaptive optimization process includes the following steps: Initialize the Aurora Optimization Algorithm population, configure the population size, maximum number of iterations and hyperparameter search range, and encode the hyperparameters of the CNN-BiGRU-Attention model; In each iteration, the fitness function value corresponding to the position of the aurora particle in the solution space is calculated, where the fitness function is the mean absolute percentage error of the CNN-BiGRU-Attention model on the validation set. In each iteration, the position of individuals in the population is updated, including global search in the exploration phase and local search in the development phase, so that the hyperparameters gradually approach the global optimum. When the iteration termination condition is met, the optimal combination of hyperparameters is output and applied to the training of the CNN-BiGRU-Attention model.

[0014] As a preferred technical solution, the updating of individual population positions is achieved using the following formula: in, , These are the individual positions before and after the update. This represents interference caused by factors such as uncontrollable environments affecting particles, and its value is [0,1]. , These are two adaptive weight values. It is a complex variation of the aurora ellipse. t Let the current iteration algebra be... T For the maximum number of iterations, For the velocity of the particle, Indicates the use of global jump Step length, Step size, It is the location of the center of mass of the high-energy particle swarm. To determine the lower bound of the solution space, To find the upper limit of the solution space, This represents interference caused by factors such as uncontrollable environments affecting particles, and its value is [0,1]. It is an indicator for adjusting stability, and its value is [0,2]. N The total number of individuals.

[0015] As a preferred technical solution, the updating of individual population locations further includes: in, , , All of these represent disturbances caused by uncontrollable environments affecting particles, and all values ​​are [0,1]. This is the proportional adjustment coefficient. a This is the adaptive adjustment coefficient.

[0016] As a preferred technical solution, the process of calculating the prediction error sequence and performing pattern recognition to obtain the pattern type includes the following steps: Calculate the predicted residual sequence : In the formula, For actual wind power output, This is a preliminary forecast of wind power output; Within the set sliding time window Within this process, feature extraction is performed on the residual sequence to obtain statistical features, including the average error. Standard deviation and the maximum single-step change ; The statistical characteristics are compared with preset thresholds, including mutation thresholds. Fluctuation threshold and system deviation threshold ; Based on the comparison results, the pattern type is obtained from the dynamic characteristics of the residual sequence. This indicates a sudden change in the residuals, classifying it as a mutant. and If the result is positive, it indicates that the residual sequence is stable and is classified as stationary; otherwise, it is classified as volatile.

[0017] As a preferred technical solution, the error correction sub-model is modeled as follows: During training, training sets are constructed for each mode type, with input feature vectors and error sequences as inputs and output power correction values ​​as outputs. The training objective is to minimize the deviation of the predicted power correction values ​​and train the corresponding LightGBM error correction sub-model. During prediction, the model is routed to the corresponding LightGBM sub-model based on the current model type. The residuals are then nonlinearly fitted using the gradient boosting decision tree structure to obtain the power correction value.

[0018] As a preferred technical solution, the final ultra-short-term wind power prediction value is obtained by summing the preliminary predicted wind power and the power correction value.

[0019] Another aspect of the present invention provides an ultra-short-term wind power prediction system based on an aurora optimization algorithm and error correction, for implementing the aforementioned ultra-short-term wind power prediction method, the system comprising: The data processing module is used to acquire historical power data of the target wind farm before the target prediction period and meteorological characteristic data of the same period, and perform normalization processing to form an input feature vector. The preliminary prediction module is used to construct a CNN-BiGRU-Attention model that takes input feature vectors as input and outputs preliminary predicted wind power. The hyperparameters of the CNN-BiGRU-Attention model are adaptively optimized based on the Aurora Optimization Algorithm. The pattern recognition module is used to calculate the prediction error sequence and perform pattern recognition based on the preliminary predicted wind power output of the optimized CNN-BiGRU-Attention model and the pre-acquired real power to obtain the pattern type. The correction module is used to call the corresponding pre-built and trained error correction sub-model based on the mode type, taking the input feature vector and error sequence as input, and outputting the power correction value; The output module is used to obtain the final ultra-short-term wind power prediction value based on the preliminary wind power prediction and the power correction value.

[0020] Compared with the prior art, the present invention has at least one of the following beneficial effects: (1) Fully modeling the nonlinear relationship between meteorological features and wind power: This invention constructs a CNN-BiGRU-Attention model that takes input feature vectors as input and preliminary predicted wind power as output. It integrates convolutional neural networks (CNN), bidirectional gated recurrent units (BiGRU), and attention mechanisms. This model can efficiently extract local key features and capture the bidirectional dependence of time series in minute-level wind power prediction tasks, and adaptively allocate weights according to feature importance. Compared with traditional single neural network models, this method can significantly improve the modeling ability of the nonlinear relationship between multidimensional meteorological features and wind power, and improve the accuracy and robustness of prediction.

[0021] (2) Improve the generalization ability of the model: Based on the Aurora optimization algorithm, this invention adaptively optimizes the hyperparameters (such as learning rate, convolution kernel size, number of BiGRU units, number of Attention heads, etc.) of the CNN-BiGRU-Attention model. By setting the minimum mean absolute percentage error (MAPE) as the objective function, the hyperparameters are updated efficiently in the multidimensional search space, thereby improving the generalization ability of the model under different wind farms and complex meteorological conditions and reducing the limitations caused by manual parameter tuning.

[0022] (3) Wind power prediction error correction is achieved: Based on the preliminary predicted wind power output of the optimized CNN-BiGRU-Attention and the pre-acquired real power, the present invention calculates the prediction error sequence and performs pattern recognition to obtain the pattern type, calls the corresponding pre-built and trained error correction sub-model, takes the input feature vector and error sequence as input, outputs the power correction value, and obtains the final ultra-short-term wind power prediction value based on the preliminary predicted wind power and the power correction value. This method significantly improves the stability and accuracy of prediction under extreme weather, low-altitude wind shear, sudden wind speed changes and other conditions. Attached Figure Description

[0023] Figure 1 This is a flowchart of the ultra-short-term wind power prediction method in the embodiment; Figure 2 This is a schematic diagram of the CNN-BiGRU-Attention model in the embodiment; Figure 3 This is a flowchart illustrating the optimization of the aurora optimization algorithm in the embodiment; Figure 4 This is a diagram illustrating the error correction mechanism in the embodiment. Detailed Implementation

[0024] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of the present invention.

[0025] Example 1 To address the problems of the aforementioned existing technologies, this embodiment provides an ultra-short-term wind power prediction method based on the Aurora Optimization Algorithm (PLO) and error correction. Based on an improved deep learning model, combined with the Aurora Optimization Algorithm (PLO) and the LightGBM error correction module, it extracts temporal features and nonlinear relationships from historical wind speed, wind direction, and power data to generate multiple possible prediction scenarios for future wind power, and then performs power prediction based on these scenarios. A pattern recognition and adaptive correction mechanism for prediction errors is introduced, further improving the accuracy and reliability of ultra-short-term wind power prediction. Through qualitative and quantitative evaluation of different prediction models, the method of this invention demonstrates significant advantages in handling wind power prediction tasks, and can better cope with the challenges brought by sudden wind speed changes and complex meteorological fluctuations.

[0026] In this embodiment, the wind power prediction based on the aurora optimization algorithm and error correction is based on the following principle: In wind power forecasting, historical power data and multi-source meteorological features (such as wind speed, wind direction, temperature, and humidity) can comprehensively reflect the fluctuation patterns of wind farm output power. To improve the accuracy of wind power forecasting under complex weather conditions, meteorological and power data should be as timely and complete as possible. Observing wind power time series data, it is easy to see that at the "minute-level" time scale, power values ​​fluctuate rapidly due to sudden changes in wind speed and wind direction. Power data at adjacent time points usually have certain correlations in local features. Therefore, a combination of convolutional neural networks (CNN) and bidirectional gated recurrent units (BiGRU) with an attention mechanism can be used to achieve feature extraction and time series modeling. To meet the requirements of speed and accuracy for minute-level power forecasting, this invention is based on a CNN-BiGRU-Attention hybrid model and introduces the Propagation Loop (PLO) algorithm to adaptively optimize the model's hyperparameters, thereby improving the model's generalization ability.

[0027] In terms of error correction, a feature extraction and pattern classification mechanism for error sequences is constructed to classify prediction errors into three states: stationary, fluctuating, and abrupt. Corresponding LightGBM correction sub-models are trained for each state. LightGBM, with its efficient gradient boosting decision tree structure, can accurately capture the nonlinear patterns in the residuals, enabling intelligent correction of preliminary prediction results. This effectively reduces prediction bias and improves the model's stability and prediction accuracy under complex weather conditions.

[0028] For details, see Figure 1 This method includes the following steps: Step S1: Obtain historical power data and various meteorological characteristic data of the wind farm during the same period, normalize the input feature values ​​to eliminate dimensional differences and improve the feature extraction effect.

[0029] In this embodiment, historical power data of wind farms and multidimensional meteorological feature data of corresponding time periods are acquired, including indicators such as wind speed, wind direction, temperature, humidity, atmospheric pressure, and air density. A multidimensional input feature vector for the prediction model is constructed, and normalization processing is used to eliminate dimensional differences between features, improving the effectiveness of subsequent feature extraction and model training. Specific implementation steps include steps S11-S13: Step S11: Obtain historical power data of the station before the target prediction period, and simultaneously collect meteorological characteristic data provided by the meteorological monitoring system, including irradiance, temperature, humidity, wind speed, wind direction and cloud cover.

[0030] Step S12: Time alignment and missing value processing are performed on the acquired historical power data and multi-source meteorological feature data to ensure the integrity and consistency of the input samples.

[0031] Step S13: The aligned input feature vector is processed by normalization to eliminate the dimensional differences between different features and improve the comparability of feature data and feature extraction effect in subsequent deep learning models.

[0032] By comprehensively utilizing historical power data and multidimensional meteorological features, and combining them with normalization processing, the comparability and usability of input features can be significantly improved, providing high-quality data support for subsequent CNN-BiGRU-Attention hybrid prediction models.

[0033] Step S2: Construct a CNN-BiGRU-Attention hybrid neural network model. CNN is used to extract local key features, BiGRU is used to capture temporal bidirectional dependencies, and the Attention mechanism is used to dynamically allocate time step weights.

[0034] See Figure 2 The CNN-BiGRU-Attention hybrid neural network model includes a coarse-connected CNN, pooling layers, flattening layers, BiGRU, self-attention mechanism layers, and fully connected layers.

[0035] In this embodiment, a CNN-BiGRU-Attention hybrid neural network prediction model is constructed, which fully combines the advantages of convolutional neural networks (CNN) in local feature extraction, the ability of bidirectional gated recurrent units (BiGRU) in temporal dependency modeling, and the role of attention mechanism in dynamic feature weighting, to achieve efficient modeling of multidimensional wind power characteristics and meteorological characteristics. The specific implementation steps include steps S21-S23: Step S21: Tensor encoding is performed on the normalized wind power and meteorological feature data in the input layer, and then input into a convolutional neural network (CNN). Multi-layer convolution and pooling operations are used to extract key local spatiotemporal features, enhancing feature representation capabilities. The CNN input consists of historical wind power data and multi-influence feature data. The data is divided into... d Every day n There are 1 data point, and each data point has 1 data point. m Each feature forms a n × m × d The matrix is ​​used as input to the CNN model. The output expression of the CNN convolutional layer is shown below: Wherein: It is an activation function. Slide window size The weights of the convolution kernel in rows n and columns m, It is the value in the nth row and mth column of the feature matrix of the input data. It is convolution kernel bias.

[0036] The CNN pooling layer uses a 3×3 filter and a sliding window with a stride of 1 for sampling, reducing the size of data features and network parameters. Then, the data is input into the BiGRU layer through a fully connected layer and a flattening layer. Step S22: The high-dimensional feature map output by the CNN is input into a Bidirectional Gated Recurrent Unit (BiGRU) network to capture temporal dependencies from both the forward and backward time dimensions, and extract bidirectional dynamic features of the input sequence. The structural expression of the GRU is shown below: in: , These are the reset door and the update door, respectively. for Input at any moment; , , This is the weight matrix; , , It is the bias vector; Use the Sigmoid activation function; For Tanh activation function; This represents an intermediate memory state, i.e., a new candidate state calculated through the reset gate; This is the current state, which is the final hidden state controlled by the updated door; This represents the output of the hidden layer neurons at the previous time step; It represents the Hadama accumulation.

[0037] BiGRU, through its bidirectional structure, can simultaneously learn past and future information from time series data. This characteristic allows BiGRU to more comprehensively understand data patterns and capture bidirectional dependencies within the sequence. Furthermore, because it considers both forward and backward information, it can more accurately understand the data's context. This improves the model's ability to understand sequence data, thereby enhancing prediction accuracy. For BiGRU, t The output at time step 1 is a weighted sum of the outputs of the forward hidden layer and the backward hidden layer, expressed as follows: in, Indicates a gated loop unit. and These are the outputs of the forward and backward hidden layers, respectively. This represents the hidden state of the forward hidden state in the previous time step. This represents the hidden state of the backward hidden state at the next time step. , They represent t The weights corresponding to the forward hidden layer state and the backward hidden layer state of BiGRU at time 1. for t The bias value corresponding to the hidden layer state at each time step; Step S23 introduces an attention mechanism into the hidden state sequence output by BiGRU, assigns weights to the importance of the predicted target based on the features at each time step, constructs a weighted feature representation, and improves the model's ability to focus on information at key moments.

[0038] An attention mechanism is introduced to dynamically adjust the importance of different global features based on the input. Instead of randomly assigning weights, a probability-based approach is used. Different features are assigned different weights, where each weight represents the importance of that feature. The principle is as follows: in, express t Attention distribution values ​​at any given time; v , W This is the weight matrix. b For bias; These are the different weight values ​​for the information in each channel of the BiGRU; Output for BiGRU The j-th component; The dimension of the input vector to the prediction model; For attention mechanisms in t Output at any given moment.

[0039] Step S3: The Aurora Optimization (PLO) algorithm is used to adaptively optimize key hyperparameters of the CNN-BiGRU-Attention model, such as the learning rate, kernel size, and number of neurons. The parameters are iteratively updated with the minimum MAPE as the objective function to improve the model's prediction accuracy and generalization ability.

[0040] In this embodiment, the Polar Lights Optimization (PLO) algorithm is used to adaptively optimize the key hyperparameters of the CNN-BiGRU-Attention hybrid neural network. By simulating the energy trajectory of aurora particles, a global optimal solution is searched, significantly improving the accuracy and robustness of the prediction model. (See [link to relevant documentation]). Figure 3 The specific implementation steps include steps S31-S34: Step S31: Initialize the Aurora Optimization Algorithm population, set the population size, maximum number of iterations and hyperparameter search range, and encode key hyperparameters such as learning rate, kernel size and number of neurons of the CNN-BiGRU-Attention model.

[0041] Step S32: In each iteration, the fitness value of the aurora particle is calculated based on its position in the solution space. The fitness function is the mean absolute percentage error (MAPE) of the model on the validation set.

[0042] Step S33: The individual positions in the population are iteratively adjusted according to the update mechanism of the aurora optimization algorithm, including global search in the exploration phase and local search in the development phase, so that the model hyperparameters gradually approach the global optimum. The individual position update formula is: in, The disturbance caused by factors such as uncontrollable environment of particles is a value of [0,1]. , These are two adaptive weight values. It is a complex variation of the aurora ellipse. and The expression is as follows: in, , It indicates that with the accumulation of time, The influence on particles is becoming increasingly significant, and at the same time Its effect on particles is becoming smaller and smaller. in, It is an important indicator for adjusting the stability of LF, with a value between 0 and 2. Step size, It is the position of the center of mass of the high-energy particle swarm, and the difference between it and the particle's current position represents the particle's tendency to move. To find the upper limit of the solution space, To determine the lower bound of the solution space, The interference caused by factors such as uncontrollable environment of particles is a value of [0,1].

[0043] Robust global and strong local strategies are only part of the PLO algorithm. To enable the algorithm to avoid convergence to local optima or escape them after getting trapped, PLO proposes a particle collision strategy. This strategy simulates chaotic collisions between the current particle and any other particle in the swarm, resulting in a random new position. As particles converge within the aurora ellipse, collisions become more frequent, causing the aurora shape to continuously change. This is achieved when the following conditions are met: At that time, the specific update model for particles is as follows: in, , , All represent the interference caused by factors such as uncontrollable environment of particles, which are values ​​of [0,1]. This means that the power of particles colliding in a chaotic manner increases over time.

[0044] Step S34: When the iteration termination condition is met, output the optimal hyperparameter combination and apply it to the training of the CNN-BiGRU-Attention model to improve prediction accuracy and generalization ability.

[0045] Step S4: The preliminary prediction results output by the optimized CNN-BiGRU-Attention model are compared with the actual power values ​​to calculate the prediction error sequence. Based on the error sequence, pattern recognition (stationary, fluctuating, and abrupt) is performed to provide a reference for subsequent error correction.

[0046] In this embodiment, the preliminary prediction results of the CNN-BiGRU-Attention hybrid neural network model optimized by the Aurora Optimization Algorithm (PLO) on the test set are compared with the corresponding actual wind power values ​​to calculate the prediction error sequence. Based on the distribution characteristics of the prediction errors, the error states are pattern-recognized and classified into three categories: stationary, fluctuating, and abrupt. By identifying different prediction states, a reference is provided for the selection and invocation of the subsequent LightGBM error correction sub-model, thereby improving the targeting and effectiveness of error correction. See also Figure 4(a) indicates that LightBGM adopts a "leaf splitting" method based on the Boosting algorithm. By selecting the leaf node with the highest information gain for splitting, the accuracy can be greatly improved. (b) indicates that LightGBM can effectively reduce memory usage and computational complexity by discretizing continuous features into a fixed number of buckets (histograms) for storage and computation.

[0047] The specific implementation steps include steps S41-S44: Step S41, calculate the predicted residual sequence, defined as: in, This is the actual power value. These are preliminary forecasts.

[0048] Step S42, within the set sliding time window Within this process, feature extraction is performed on the residual sequence, and calculations are made including the average error. Standard deviation and the maximum single-step change .

[0049] Step S43: Compare the above statistical features with a preset threshold and take historical data. The 95th percentile of the distribution is the mutation threshold. Take history The 40th percentile of the distribution is the fluctuation threshold. Take history The median of the distribution is the systematic bias threshold. .

[0050] Step S44: Based on the comparison results, perform pattern segmentation on the dynamic characteristics of the residual sequence: If If so, it indicates a sudden change in the residual, and it is determined to be a mutant type; if and If the result is positive, it indicates that the residual sequence is stable and is classified as stationary; otherwise, it is classified as volatile.

[0051] Step S5: For different prediction states, call the corresponding LightGBM error correction sub-model, use the gradient boosting decision tree structure of LightGBM to fit the prediction residual, capture complex nonlinear relationships, and output the correction value.

[0052] In this embodiment, based on the predicted state (stationary, fluctuating, or abrupt) identified in step S4, the corresponding LightGBM error correction sub-model is invoked to perform residual fitting and correction on the preliminary prediction results output by the CNN-BiGRU-Attention model. Specific implementation steps include steps S51-S53: Step S51: Based on the predicted state classification results obtained in S4, determine the LightGBM sub-model to be invoked.

[0053] Step S52: Construct training sets for each predicted state, use the predicted residuals as training targets, and use the original input features and error information as model inputs to train the corresponding LightGBM error correction sub-model.

[0054] Step S53: In actual prediction, the model is routed to the corresponding LightGBM sub-model according to the current prediction state, and its gradient boosting decision tree structure is used to perform nonlinear fitting on the residuals to obtain the correction value.

[0055] Step S6: Sum the preliminary prediction results with the error correction value to obtain the final wind power prediction result.

[0056] In this embodiment, the preliminary prediction result output by the optimized CNN-BiGRU-Attention hybrid neural network model is summed with the correction value output by the corresponding LightGBM error correction sub-model to obtain the final wind power prediction result. By introducing an error correction mechanism for different prediction states on the basis of the main prediction model, the prediction bias can be effectively reduced, the overall prediction accuracy and the model's generalization ability can be improved, and it is especially suitable for minute-level ultra-short-term wind power prediction tasks. The specific implementation steps include steps S61-S62: Step S61: Receive the preliminary prediction result output by the CNN-BiGRU-Attention model and the residual correction value output by the LightGBM error correction sub-model, the correction value corresponding to the prediction state determined in S4.

[0057] Step S62: Sum the preliminary prediction results and the residual correction values ​​to obtain the corrected final wind power prediction value.

[0058] In summary, this embodiment constructs a hybrid deep prediction model that integrates convolutional neural networks (CNN), bidirectional gated recurrent units (BiGRU), and attention mechanisms. This model efficiently extracts key local features and captures bidirectional time-series dependencies in minute-level wind power prediction tasks, adaptively allocating weights based on feature importance. Compared to traditional single neural network models, this method significantly improves the ability to model the nonlinear relationship between multidimensional meteorological features and wind power, enhancing prediction accuracy and robustness.

[0059] Furthermore, this method introduces the Aurora Optimization algorithm to adaptively optimize key hyperparameters (such as learning rate, kernel size, number of BiGRU units, and number of attention heads) in the CNN-BiGRU-Attention hybrid model. By setting the minimum mean absolute percentage error (MAPE) as the objective function, the hyperparameters are updated efficiently in a multi-dimensional search space, thereby improving the model's generalization ability under different wind farms and complex weather conditions and reducing the limitations of manual parameter tuning.

[0060] Finally, this embodiment provides a LightGBM error correction mechanism based on prediction error pattern recognition. First, based on the residual between the optimized master prediction model output and the true value, pattern classification is performed using three prediction states: stationary, fluctuating, and abrupt. Then, LightGBM sub-models are trained for different error patterns to achieve fitting and compensation for complex nonlinear residuals. This method significantly improves the stability and accuracy of predictions under extreme weather conditions, low-level wind shear, and sudden wind speed changes.

[0061] Example 2 Based on Example 1, this example provides an ultra-short-term wind power prediction system based on the aurora optimization algorithm and error correction, used to implement the ultra-short-term wind power prediction method of Example 1. The system includes: (1) Data processing module, used to acquire historical power data of the target wind farm before the target prediction period and meteorological characteristic data of the same period, and perform normalization processing to form input feature vector.

[0062] (2) Preliminary prediction module, used to construct a CNN-BiGRU-Attention model with input feature vector and output of preliminary wind power, and to adaptively optimize the hyperparameters of the CNN-BiGRU-Attention model based on the Aurora optimization algorithm.

[0063] (3) Pattern recognition module, which is used to calculate the prediction error sequence and perform pattern recognition based on the preliminary predicted wind power output of the optimized CNN-BiGRU-Attention and the pre-acquired real power to obtain the pattern type.

[0064] (4) Correction module, which is used to call the corresponding pre-built and trained error correction sub-model based on the mode type, and output the power correction value with the input feature vector and error sequence as input.

[0065] (5) Output module, used to obtain the final ultra-short-term wind power prediction value based on the preliminary wind power prediction and power correction value.

[0066] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any person skilled in the art can easily conceive of various equivalent modifications or substitutions within the technical scope disclosed in the present invention, and these modifications or substitutions should all be covered within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A method for ultra-short-term wind power prediction based on aurora optimization algorithm and error correction, characterized in that, Includes the following steps: Historical power data of the target wind farm before the target prediction period and meteorological characteristic data of the same period are obtained and normalized to form an input feature vector; Construct a CNN-BiGRU-Attention model that takes input feature vectors as input and outputs preliminary predicted wind power; The hyperparameters of the CNN-BiGRU-Attention model are adaptively optimized based on the Aurora Optimization algorithm; Based on the preliminary predicted wind power output of the optimized CNN-BiGRU-Attention model and the pre-acquired real power, the prediction error sequence is calculated and pattern recognition is performed to obtain the pattern type. Based on the aforementioned mode type, the corresponding pre-built and trained error correction sub-model is invoked, taking the input feature vector and error sequence as input, and outputting the power correction value; Based on the preliminary wind power forecast and the power correction value, the final ultra-short-term wind power forecast value is obtained.

2. The ultra-short-term wind power prediction method based on aurora optimization algorithm and error correction according to claim 1, characterized in that, The process of obtaining the input feature vector includes the following steps: Historical power data of the target wind farm before the target prediction period and meteorological characteristic data of the same period are obtained, wherein the meteorological characteristics include irradiance, temperature, humidity, wind speed, wind direction and cloud cover; Historical power data and meteorological feature data are time-aligned and missing values ​​are processed to form an input feature vector; The input feature vector is normalized.

3. The ultra-short-term wind power prediction method based on aurora optimization algorithm and error correction according to claim 1, characterized in that, The CNN-BiGRU-Attention model is modeled as follows: The input feature vector is tensor-encoded and fed into the built-in convolutional neural network, which extracts local spatiotemporal key features through multi-layer convolution and pooling operations. The high-dimensional feature maps output by the built-in convolutional neural network are input into the built-in bidirectional gated recurrent unit network to capture temporal dependencies from both forward and backward time dimensions and extract bidirectional dynamic features of the input sequence. Attention processing is performed on the hidden state sequence output by the bidirectional gated recurrent unit network, and weights are assigned to the importance of the predicted target according to the features at each time step to construct a weighted feature representation; The preliminary predicted wind power output is based on the weighted feature representation.

4. The ultra-short-term wind power prediction method based on aurora optimization algorithm and error correction according to claim 1, characterized in that, The adaptive optimization process includes the following steps: Initialize the Aurora Optimization Algorithm population, configure the population size, maximum number of iterations and hyperparameter search range, and encode the hyperparameters of the CNN-BiGRU-Attention model; In each iteration, the fitness function value corresponding to the position of the aurora particle in the solution space is calculated, where the fitness function is the mean absolute percentage error of the CNN-BiGRU-Attention model on the validation set. In each iteration, the position of individuals in the population is updated, including global search in the exploration phase and local search in the development phase, so that the hyperparameters gradually approach the global optimum. When the iteration termination condition is met, the optimal combination of hyperparameters is output and applied to the training of the CNN-BiGRU-Attention model.

5. The ultra-short-term wind power prediction method based on aurora optimization algorithm and error correction according to claim 4, characterized in that, The update of individual population positions is achieved using the following formula: in, , These are the individual positions before and after the update. This represents interference caused by factors such as uncontrollable environments affecting particles, and its value is [0,1]. , These are two adaptive weight values. It is a complex variation of the aurora ellipse. t Let the current iteration algebra be... T For the maximum number of iterations, For the velocity of the particle, Indicates the use of global jump Step length, Step size, It is the location of the center of mass of the high-energy particle swarm. To determine the lower bound of the solution space, To find the upper limit of the solution space, This represents interference caused by factors such as uncontrollable environments affecting particles, and its value is [0,1]. It is an indicator for adjusting stability, and its value is [0,2]. N The total number of individuals.

6. The ultra-short-term wind power prediction method based on aurora optimization algorithm and error correction according to claim 5, characterized in that, The update of individual population locations also includes: in, , , All of these represent disturbances caused by uncontrollable environments affecting particles, and all values ​​are [0,1]. This is the proportional adjustment coefficient. a This is the adaptive adjustment coefficient.

7. The ultra-short-term wind power prediction method based on aurora optimization algorithm and error correction according to claim 1, characterized in that, The process of calculating the prediction error sequence and performing pattern recognition to obtain the pattern type includes the following steps: Calculate the predicted residual sequence : In the formula, For actual wind power output, This is a preliminary forecast of wind power output; Within the set sliding time window Within this process, feature extraction is performed on the residual sequence to obtain statistical features, including the average error. Standard deviation and the maximum single-step change ; The statistical characteristics are compared with preset thresholds, including mutation thresholds. Fluctuation threshold and system deviation threshold ; Based on the comparison results, the pattern type is obtained from the dynamic characteristics of the residual sequence. This indicates a sudden change in the residuals, classifying it as a mutant. and If the result is positive, it indicates that the residual sequence is stable and is classified as stationary; otherwise, it is classified as volatile.

8. The ultra-short-term wind power prediction method based on aurora optimization algorithm and error correction according to claim 1, characterized in that, The error correction sub-model is modeled as follows: During training, training sets are constructed for each mode type, with input feature vectors and error sequences as inputs and output power correction values ​​as outputs. The training objective is to minimize the deviation of the predicted power correction values ​​and train the corresponding LightGBM error correction sub-model. During prediction, the model is routed to the corresponding LightGBM sub-model based on the current model type. The residuals are then nonlinearly fitted using the gradient boosting decision tree structure to obtain the power correction value.

9. The ultra-short-term wind power prediction method based on aurora optimization algorithm and error correction according to claim 1, characterized in that, By summing the preliminary wind power forecast and the power correction value, the final ultra-short-term wind power forecast is obtained.

10. An ultra-short-term wind power prediction system based on aurora optimization algorithm and error correction, characterized in that, For implementing the ultra-short-term wind power prediction method as described in any one of claims 1-9, the system comprises: The data processing module is used to acquire historical power data of the target wind farm before the target prediction period and meteorological characteristic data of the same period, and perform normalization processing to form an input feature vector. The preliminary prediction module is used to construct a CNN-BiGRU-Attention model that takes input feature vectors as input and outputs preliminary predicted wind power. The hyperparameters of the CNN-BiGRU-Attention model are adaptively optimized based on the Aurora Optimization Algorithm. The pattern recognition module is used to calculate the prediction error sequence and perform pattern recognition based on the preliminary predicted wind power output of the optimized CNN-BiGRU-Attention model and the pre-acquired real power to obtain the pattern type. The correction module is used to call the corresponding pre-built and trained error correction sub-model based on the mode type, taking the input feature vector and error sequence as input, and outputting the power correction value; The output module is used to obtain the final ultra-short-term wind power prediction value based on the preliminary wind power prediction and the power correction value.

Images