Wind power prediction system for optimizing neural network based on genetic algorithm

Abstract

The invention discloses a wind power prediction system for optimizing a neural network based on a genetic algorithm, relates to the technical field of new energy power system prediction, and improves the precision and adaptability of wind power prediction by fusing the genetic algorithm and a deep neural network. The system adopts multi-objective genetic optimization, randomly initializes a neural network parameter combination, evaluates the fitness by taking a prediction error and model complexity as double objectives, and screens out an optimal network architecture through evolution operation; in the aspect of neural network training, the system adopts an LSTM and TCN hybrid network as a basic model, dynamic weighting input features of a meteorological attention mechanism are combined, a learning rate and regularization parameters are optimized by using a genetic algorithm, model convergence is accelerated, and overfitting is prevented; in addition, for the space-time imbalance of the wind power data, a generative adversarial network is introduced to generate synthetic data in an extreme weather scene, and the generalization ability of the model is enhanced.

Description

Technical Field

[0001] The present invention relates to the technical field of new energy power data analysis and prediction, and in particular to a wind power prediction system based on a genetic algorithm optimized neural network. Background Art

[0002] As a clean, renewable energy source, wind power has become a crucial component of the global energy transition. However, due to the complex influence of meteorological factors such as wind speed, direction, temperature, humidity, and air pressure, wind power exhibits significant nonlinearity, randomness, and uncertainty. This makes accurate wind power forecasting a key challenge in wind power grid integration and operation scheduling. The accuracy of wind power forecasting is directly linked to the safe and stable operation of the power grid and the economic benefits of wind farms. Therefore, improving the accuracy and robustness of wind power forecasting has become a key research and engineering practice issue.

[0003] At present, the common wind power prediction methods on the market are mainly divided into the following categories: 1. Traditional physical modeling methods: These methods predict wind power by building mathematical models based on wind farm topography, meteorological conditions, and wind turbine characteristics. Typical methods include meteorological dynamic models and numerical weather prediction models. These methods perform well in predicting wind power trends over longer time periods (e.g., hours to days). However, due to the complexity of the models, the high computational effort, and their strong dependence on the site environment and equipment characteristics, their accuracy is often significantly limited. This makes them particularly difficult to meet practical needs in short-term and ultra-short-term forecasts.

[0004] 2. Statistical-based methods: Statistical methods analyze the correlation between historical data and meteorological variables, using tools such as time series analysis, linear regression, and support vector machines (SVM) to forecast wind power. These methods are computationally simple and suitable for short-term forecasts. However, due to the highly nonlinear and multidimensional correlations in wind power data, traditional linear statistical models are prone to large prediction errors when processing complex data and have difficulty capturing the power fluctuations under extreme weather conditions.

[0005] 3. Artificial Intelligence-based methods: With the development of machine learning and deep learning technologies, more and more researchers are using techniques such as neural networks, long short-term memory networks (LSTMs), and convolutional neural networks (CNNs) to build wind power prediction models. These methods offer significant advantages when processing nonlinear and multidimensional time series data, significantly improving prediction accuracy. However, traditional neural network models rely on randomly initialized weights and thresholds during training, making them prone to local optima. The selection of model hyperparameters is complex, making it difficult to ensure prediction accuracy and stability.

[0006] The limitations of existing technologies are mainly reflected in the following aspects: 1. Local optimality problem: Traditional neural networks are prone to falling into local optimal solutions during model training and are unable to fully explore the global patterns in the data, resulting in large prediction errors.

[0007] 2. Insufficient model generalization ability: Due to the strong randomness and uncertainty of wind power data, the existing methods have weak model generalization ability when facing complex environments and extreme meteorological conditions, and it is difficult to provide stable and reliable prediction results.

[0008] 3. Difficulty in selecting hyperparameters: The hyperparameters of neural networks (such as the number of network layers, number of nodes, learning rate, etc.) have a significant impact on model performance, but most existing methods rely on manual experience or random selection, resulting in large fluctuations in model performance.

[0009] 4. Limited real-time and adaptive capabilities: Wind power forecasting requires processing a large amount of data collected in real time. Existing models are difficult to achieve rapid updates and adaptive adjustments during the forecasting process, resulting in the forecast results lagging behind actual changes.

[0010] To address the above issues, the present invention proposes a wind power prediction system based on a genetic algorithm-optimized neural network. As a global optimization method, the genetic algorithm can efficiently find the optimal parameter combination within the global search space by simulating the evolutionary process (selection, crossover, and mutation) in nature, avoiding the drawbacks of traditional gradient descent methods that are prone to falling into local optimality. By globally optimizing the weights, thresholds, and hyperparameters of the neural network through the genetic algorithm, the present invention significantly improves the accuracy and generalization capabilities of the wind power prediction model. Furthermore, the system possesses adaptive learning and online update capabilities, capable of continuously optimizing the prediction model based on real-time data, ensuring the stability and real-time nature of the prediction results, thereby providing more accurate decision support for wind farm operation and scheduling. Summary of the Invention

[0011] In view of the shortcomings of the existing technology, the present invention provides a wind power prediction system based on a genetic algorithm optimized neural network, which solves the problem of insufficient accuracy of wind power prediction.

[0012] To achieve the above objectives, the present invention is implemented through the following technical solutions: a wind power prediction system based on a genetic algorithm-optimized neural network. By introducing a genetic algorithm (GA) to optimize the parameters and structure of the neural network model, the model's prediction accuracy is significantly improved under nonlinear and multidimensional data. By simulating the processes of natural selection, crossover, and mutation, the genetic algorithm can efficiently find the optimal solution within the global search space, avoiding being trapped in a local optimum. Specific implementations include: Encode the weights and thresholds of the neural network, construct a fitness function, and use the prediction error as the optimization target; The global optimization of neural network parameters is achieved by dynamically adjusting the parameter combination through the iterative process of genetic algorithm; The optimized network structure is tested under different meteorological conditions and data sets to ensure that the model has strong adaptability and robustness.

[0013] This global optimization mechanism enables the prediction model to better capture the complex patterns and underlying laws in wind power variations, thereby improving the overall prediction accuracy.

[0014] Avoid local optimum: use genetic algorithm to search parameter space globally; Genetic algorithms utilize a swarm search strategy to simultaneously explore multiple possible parameter combinations, avoiding the problem of gradient descent methods relying solely on local gradient information for updates. By optimizing the weights, thresholds, and structural configuration of a neural network using genetic algorithms, the risk of the model falling into a local optimal solution can be effectively avoided.

[0015] In the initial stage, multiple candidate solutions (i.e., neural network parameter combinations) are randomly generated to ensure the diversity of the search space; In each iteration, a new generation of parameter combinations is generated by selecting solutions with higher fitness for crossover and mutation; Repeat the iteration until the fitness converges, and finally obtain the neural network model with the best performance.

[0016] This method can significantly improve the convergence and robustness of the model, and enhance prediction accuracy and stability.

[0017] Automated hyperparameter optimization: using genetic algorithms to select the optimal hyperparameter combination; This paper uses a genetic algorithm to automatically optimize the hyperparameters of a neural network (such as the number of hidden layers, the number of nodes per layer, the learning rate, the type of activation function, etc.), replacing traditional manual selection or grid search. The specific steps are as follows: Encoding hyperparameters as chromosomes of genetic algorithms; Construct a fitness function based on cross-validation error to evaluate the performance of each set of hyperparameter combinations; Automatically find the optimal hyperparameter combination through the selection, crossover and mutation process of the genetic algorithm.

[0018] This method greatly improves the degree of automation of the model, avoids the uncertainty caused by manual selection of parameters, and ensures the stability and reliability of the model performance.

[0019] Enhance real-time and adaptive capabilities: introduce online learning and dynamic update mechanisms; To meet the needs of real-time prediction, the present invention integrates online learning and adaptive update functions into the system, which can dynamically adjust the neural network parameters according to the new data collected in real time, thereby improving the long-term stability and prediction accuracy of the model. The specific implementation includes: While the model is predicting, the deviation between the actual wind power and the predicted results is continuously monitored; When the deviation exceeds the set threshold, the model's adaptive update is triggered, and the model is incrementally trained using the latest data; The sliding window technology is used to limit the amount of updated data to prevent the model from overfitting due to excessive real-time data.

[0020] Data preprocessing and outlier processing: ensure data quality; To address data missing and anomaly issues, this paper designs an efficient data preprocessing module and uses the following methods to improve data quality: Outlier detection: Use the KNN algorithm based on density peaks to screen outliers in real-time collected data and eliminate noise data; Missing value filling: Bayesian inference method is used to interpolate missing values to restore data integrity to the greatest extent; Data normalization and feature extraction: Normalize the processed data and extract key features to ensure that the neural network can efficiently use the data for prediction.

[0021] The present invention provides a wind power prediction system based on a genetic algorithm-optimized neural network. Compared with the existing technology, it has the following advantages: This paper comprehensively considers issues such as insufficient data under extreme weather conditions, limited meteorological feature screening and data scarcity, insufficient model generalization, and complex time series data processing. It proposes a wind power forecasting system based on a genetic algorithm-optimized neural network. Through various technological innovations, including a combined confidence index, transfer learning, multi-source data fusion, and an improved neural network algorithm, this system significantly improves the accuracy and stability of wind power forecasting, providing a new solution for forecasting technology in the new energy sector.

[0022] Compared with the existing wind power prediction technology on the market, the present invention significantly improves the prediction accuracy, model stability and real-time adaptability, solves the limitations faced by traditional methods in processing nonlinear, multi-dimensional complex data, and brings about many technical effects.

[0023] First, by optimizing the neural network through a genetic algorithm, this invention effectively overcomes the vulnerability of traditional neural networks to local optima. This allows for global search and optimization of network weights, thresholds, and hyperparameters, improving the overall performance of the prediction model. The genetic algorithm comprehensively explores the parameter space, ensuring that the prediction model more accurately captures the complex relationship between wind power and various meteorological factors, reducing prediction errors and significantly improving the accuracy of short-term and ultra-short-term forecasts.

[0024] Secondly, the present invention possesses adaptive learning and dynamic updating capabilities, continuously optimizing the neural network model based on new data collected in real time, ensuring that the model's predictions always reflect the latest wind farm operating conditions. This online learning mechanism enables the system to rapidly respond to changing meteorological conditions, improving the model's real-time performance and long-term stability, and preventing predictions from lagging behind actual conditions.

[0025] In addition, the system introduces a density-peak-based KNN algorithm and Bayesian inference methods for data preprocessing and outlier detection, effectively improving data quality. By removing outliers and properly filling in missing values, the accuracy and completeness of the model training data are ensured, further improving prediction results and enhancing the robustness of the model.

[0026] Finally, this invention supports visualization of prediction results and multidimensional data analysis, enabling wind farm operators to monitor wind power trends and errors in real time, aiding in informed decision-making. This technology not only improves the safety and economic efficiency of wind farm operations but also provides technical support for large-scale wind power grid integration, demonstrating broad market application value and potential for widespread adoption. BRIEF DESCRIPTION OF THE DRAWINGS

[0027] Figure 1 This is a schematic diagram of the principle framework of the present invention; Figure 2 It is the data flow chart of genetic algorithm optimized neural network power prediction. DETAILED DESCRIPTION

[0028] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0029] Combine Figure 1 as well as Figure 2The wind power prediction system of the present invention is composed of multiple modules, including a data acquisition module, a data preprocessing module, a genetic algorithm optimization module, a neural network training module, a prediction and output module, and an adaptive update module. The modules work together to achieve accurate prediction and real-time adjustment of wind power.

[0030] The implementation of the system follows these steps: The data acquisition module is responsible for acquiring historical and real-time data from the wind farm monitoring system, including multi-dimensional meteorological and power data such as wind speed, wind direction, temperature, humidity, air pressure, and wind power output. These data come from a wide range of sources, including data from the wind farm's own monitoring equipment and meteorological forecast data provided by the meteorological department.

[0031] In order to improve the timeliness and integrity of data, the data acquisition module adopts a distributed data acquisition framework, which can collect data on different nodes and upload it to the central database; the collected data will be marked according to the timestamp for subsequent processing and analysis.

[0032] The data preprocessing module preprocesses the data set collected within a specified period and gradually identifies the abnormal values, missing values or noise in the data set: Prioritize outlier confirmation for the same type of data collected by different nodes at the same time in the data set, confirm the total number of different types of data N, and confirm the corresponding type of data x i The local density associated with the sample point is: Confirm the local density P j , where χ(……) is a unit step function, when (d ij -dc) < 0, the value is 1, otherwise it is 0, d ij Represents the corresponding type data x i with x j The distance between them (which can also be understood as the difference, where the difference is greater than 0), where dc is the cutoff distance, is obtained by: , where argmin is the abbreviation of argument of the minimum, which means taking the minimum value, and d is the d value that needs to be solved, that is, dc; Use: (P j ×C1)÷(d ij × C2) = Z j Determine the comprehensive evaluation value Z j and the comprehensive evaluation value Z j Compare and check with the preset threshold: Will satisfy: Z jThe value ≤Y1 is calibrated as an abnormal value, otherwise no calibration is performed. Y1 is a preset value, which is prepared by the operator in advance based on experience. C1 and C2 are both preset fixed coefficient factors, and their specific values are prepared by the operator based on experience; For the determined x i , extract other multi-dimensional data associated with this data, and for each set of multi-dimensional data and this x i Perform correlation calculations on the Pearson correlation coefficient to obtain a determination value. If there is a set of determination values that satisfies the following condition: determination value ≥ 0.7 (threshold), it means that the outlier is related to other data and may be valid data, so it should be retained. Otherwise, it should be eliminated.

[0033] Then, the missing value interpolation is performed on the data set after the outliers are removed. This method infers the most likely missing value based on the probability distribution of the existing data to ensure the continuity and rationality of the data: Assume that a single data follows a multivariate normal distribution N(μ,Σ), where μ is the mean vector and Σ is the covariance matrix. The data is divided into the observed part Xo and the missing part Xm, forming a data vector = (Xo,Xm); Based on the observed data Xo, the posterior distribution of the distribution parameters μ and Σ is calculated using the Bayesian estimation method. For example, for the mean vector μ, its conjugate prior distribution is a normal distribution. After combining the observed data, the posterior distribution is still a normal distribution, and its mean and covariance are updated using a specific formula. Based on the known posterior distribution of parameters, according to the properties of multivariate normal distribution, calculate the conditional distribution P(Xm|Xo) of the missing value Xm under the given Xo condition; this conditional distribution is also a normal distribution, and its mean and covariance can be derived from Xo, μ and Σ; The conditional distribution P(Xm|Xo) of the missing values is sampled to obtain multiple possible interpolation values. To reduce interpolation bias, the mean or median of these sampled values is usually selected as the final missing value interpolation result.

[0034] Then, all input data is normalized to eliminate the magnitude differences between different data dimensions and improve the convergence speed of the neural network. At the same time, key features are extracted to construct training sets and validation sets for subsequent model training: Normalize all data of the same type in a dataset and map them to the [0, 1] interval. This is suitable for scenarios where there are clear requirements for the data range and there are no extreme outliers. For example, in image data processing, normalizing pixel values to the [0, 1] interval facilitates subsequent calculations. The original data is converted into a set of new unrelated features (principal components) through linear transformation. These principal components are sorted according to the size of the variance. The larger the variance, the more information it contains. During the calculation process, the data is first centered (minus the mean), and then the covariance matrix is calculated. The eigenvectors and eigenvalues are obtained through eigenvalue decomposition. The first k eigenvectors with larger eigenvalues are selected to construct a projection matrix, and the original data is projected into a new feature space to achieve dimensionality reduction. For example, when processing high-dimensional image data, PCA can reduce thousands of dimensions of features to dozens of dimensions while retaining most of the information. For example, recursive feature elimination (RFE) is a model-based feature importance assessment that starts from the full feature set and eliminates the features that contribute the least to the model performance each time until the preset number of features is reached. The chi-square test selects features with high correlation with the target variable by calculating the independence between the features and the target variable. After feature extraction, the processed data is divided into a training set and a validation set in a certain ratio (such as 7:3 or 8:2). The training set is used for model parameter learning, allowing the model to learn the relationship between features and target variables from the data. The validation set is used to evaluate model performance during model training and adjust hyperparameters (such as the number of neural network layers, learning rate, etc.) to avoid model overfitting and ensure that the model has good generalization ability in practical applications.

[0035] Among them, the genetic algorithm optimization module optimizes the initial weights, thresholds and network structure of the neural network to avoid the problem that traditional methods are prone to fall into local optimality. The detailed optimization processing steps are as follows: Step 1: Initialize the population: Use real number encoding to directly map the weights, thresholds, and hyperparameters of the neural network to chromosomes. Take a simple three-layer neural network (input layer-hidden layer-output layer) as an example. Assume that there are m neuron connections from the input layer to the hidden layer, n neuron connections from the hidden layer to the output layer, p thresholds in the hidden layer, q thresholds in the output layer, and r hyperparameters such as the learning rate. Then the length of a chromosome is m+n+p+q+r, and each gene position corresponds to a weight, threshold, or hyperparameter value. This encoding method avoids the decoding complexity of binary encoding and is more suitable for optimizing continuous parameters. Set the population size to G, and create G chromosomes by randomly generating values within a reasonable range to form the initial population. For example, weights and thresholds can be randomly selected in the interval [−1, 1]. Hyperparameters such as the learning rate are randomly generated within an appropriate interval (e.g., [0.001, 0.1]) based on experience. To ensure population diversity, various random strategies can be adopted, such as changing the seed for random number generation or randomly selecting values in different regions to prevent the initial population from being too concentrated in a single area. Step 2: Design the fitness function: Use the prediction error as the core of the fitness function and use the mean square error (MSE) for accuracy prediction. MSE can intuitively reflect the degree of deviation between the model prediction results and the real data. The smaller the MSE, the more accurate the model prediction. To prevent model overfitting, a regularization term is introduced into the fitness function. L2 regularization (weight decay) is used. At the same time, to evaluate the generalization ability of the model, a cross-validation method is used to divide the dataset into a training set and a validation set. The performance on the validation set is used as part of the fitness function. The performance of the model on different data is comprehensively considered, so that the optimized model has better generalization effect in practical applications. Step 3: Calculate the proportion of each individual's fitness to the total fitness as the probability of being selected. Using a roulette wheel, randomly select individuals to enter the next generation population based on the probability. For example, a disk is divided into areas according to the probability of each individual. The disk is rotated, and the individual corresponding to the area where the pointer stops is selected. This method is simple and intuitive, but there is a possibility that individuals with high fitness will be over-selected, while individuals with low fitness will be ignored. Randomly select two chromosomes as parents and set the crossover point (for real-number coding, one or more gene positions can be selected as the crossover point). For example, select two crossover points and exchange the gene segments between the two chromosomes at the crossover points to generate two new daughter chromosomes. Assuming the parent chromosomes A = [a1, a2, a3, a4, a5] and B = [b1, b2, b3, b4, b5], and the crossover points are at the 2nd and 4th positions, then the daughter chromosomes C = [a1, b2, b3, a4, a5] and D = [b1, a2, a3, b4, b5]. The crossover probability is generally set to 0.6-0.9. A higher crossover probability can speed up the convergence of the algorithm, but too high a crossover probability may cause the genes of excellent individuals to be destroyed. Each gene position of an individual is mutated with a pre-set mutation probability Pm (usually between 0.001 and 0.01). If a gene position is selected for mutation, a new value is randomly generated within the range of that position's values, replacing the original gene value. For example, a gene position with an original weight of 0.3 may become -0.2 after mutation. Mutation can increase population diversity, prevent the algorithm from prematurely converging to a local optimal solution, and allow the algorithm to explore a wider solution space during the search process. Step 4: Set a maximum number of evolutionary generations (e.g., 500). The genetic algorithm terminates when this number is reached. Alternatively, the algorithm is considered converged and terminated when the optimal fitness value of the population no longer significantly improves (the increase is less than a set threshold, such as 0.001) over multiple generations (e.g., 50 generations). The neural network weights, thresholds, and hyperparameters corresponding to the optimal individual are the optimized results, which can be used to initialize the neural network for subsequent training and prediction tasks.

[0036] The neural network training module uses a multi-layer perceptron (MLP) or a long short-term memory network (LSTM) as a prediction model and uses parameters optimized by a genetic algorithm for model training. The specific training process is as follows: Step 1: Model selection and construction: The LSTM network model is selected. For data with time series characteristics, the LSTM network has strong modeling capabilities, can capture long-term dependencies and reduce information loss: The input layer receives preprocessed time series data, which needs to be converted into a three-dimensional tensor in the format of (number of samples, time steps, number of features). Stacking multiple LSTM layers can enhance the model's ability to learn complex temporal dependencies, but too many layers will significantly increase training complexity and the risk of overfitting. Generally, 2-3 layers are more appropriate. The number of LSTM neurons in each layer needs to be fine-tuned based on the data size and task difficulty. For scenarios with small data volumes and simpler tasks, 64 neurons can be used; if the data volume is large and the relationships are complex, 128 or 256 neurons can be used. Information exchange between LSTM layers is achieved through state transfer. The hidden state and cell state of the previous layer serve as the input of the next layer to further explore the temporal characteristics in the data. The number of neurons and activation function in the output layer are determined by the task type. For regression tasks (such as predicting numerical results), the output layer uses one neuron and a linear activation function to ensure that the output is a continuous value. For classification tasks (such as determining whether a device is operating normally), the number of neurons is determined based on the number of categories. For example, for a classification task with two categories, "normal" and "abnormal," the output layer has two neurons and uses a Softmax activation function to convert the output into a probability distribution for each category.

[0037] Step 2: Training process: The preprocessed training set is fed into the neural network for multiple iterations of training: During the forward propagation of the LSTM, data is fed into the network sequentially, in time steps. At each time step, the input gate determines the degree of retention of the current input information, the forget gate controls whether to discard or retain historical information in the cell state, and the output gate generates the current hidden state based on the cell state. The outputs of these gates are transformed using nonlinear activation functions (such as the sigmoid function for gated output and the tanh function for processing cell state) to ensure efficient transmission and processing of information. Use the validation set to evaluate the performance of the model, and improve the training effect by adjusting the learning rate, adaptive weight decay, and early stopping mechanism to prevent overfitting.

[0038] Step 3: Model evaluation: Indicators such as mean square error (MSE), root mean square error (RMSE) and mean absolute error (MAE) are used to comprehensively evaluate model performance to ensure that the model meets the prediction accuracy requirements.

[0039] Among them, the prediction and output module is used to use the trained neural network model to make real-time predictions and display the prediction results in a visual form or output them to the wind farm dispatching system: (1) Real-time prediction Real-time meteorological data is collected as input, and a neural network model is used to predict the wind power trend within a certain period of time in the future. The prediction period can be flexibly adjusted according to actual needs, supporting short-term prediction (1-6 hours) and ultra-short-term prediction (15 minutes to 1 hour): Real-time meteorological data collection, including but not limited to wind speed, wind direction, temperature, air pressure, and humidity, is performed using meteorological monitoring equipment (e.g., anemometers, wind vanes, barometers, and thermometers). The collection frequency is flexibly set based on the forecast period. For example, for ultra-short-term forecasts (15 minutes to 1 hour), data collection may be performed every second or every minute, while for short-term forecasts (1-6 hours), data collection may be performed every 5-10 minutes. The collected data must be pre-processed in real time. The pre-processed real-time meteorological data is organized into a three-dimensional tensor (number of samples, time steps, number of features) that conforms to the LSTM model input format and input into the trained neural network model for inference. Set the prediction period according to actual needs: Ultra-short-term forecasting: For forecasting needs of 15 minutes to 1 hour, the model is used to continuously predict wind power for multiple time steps in the future (such as every 15 minutes as a time step), providing instant reference for real-time power control of wind farms and grid scheduling. Short-term forecasts: For forecasts of 1-6 hours, the model outputs hourly wind power trends, helping operations personnel plan equipment maintenance and power generation schedules in advance. During the forecasting process, a rolling forecast approach can be used. After each forecast, newly collected real-time data is added to the dataset, the earliest data is discarded, and the input tensor is reconstructed for the next forecast to adapt to dynamic changes in meteorological conditions.

[0040] (2) Result display and alarm function Forecast results are displayed through a visual interface, including power forecast curves, error analysis charts, and meteorological parameter trend charts, enabling operators to monitor the wind farm in real time. Furthermore, the system features an alarm function that automatically triggers warnings when forecast results exceed set thresholds, enabling operators to adjust wind farm operation strategies in a timely manner.

[0041] Among them, the adaptive update module, in order to improve the long-term accuracy and stability of the model, the system has designed an adaptive update module that can dynamically adjust the neural network parameters according to real-time data to ensure that the prediction model always maintains the optimal state.

[0042] (1) Data Incremental Learning After each prediction, the newly collected data is combined with the existing data to form an incremental data set. The sliding window technology is used to control the data size and prevent the model from overfitting due to excessive data: After each forecast is completed, the newly collected meteorological data and actual wind power data are integrated with the existing data to form an incremental data set. The new data includes real-time meteorological parameters such as wind speed, wind direction, temperature, air pressure, humidity, and the actual wind power value of the corresponding time period. For example, after completing an ultra-short-term forecast, the newly collected 15-minute data for the next period is included in the data set; To avoid excessive data size leading to decreased model training efficiency or overfitting, a sliding window technique is used to manage the dataset: Window Size Setting: Determine the sliding window size based on forecast requirements and data characteristics. For very short-term forecasts, the window can include the most recent 1-2 hours of data; for short-term forecasts, it can be set to include the most recent 6-12 hours of data. For example, setting the window size to 100 data samples, each sample corresponds to 1 minute of meteorological and power data. Data update strategy: Whenever new data is added, the oldest corresponding number of data is removed from the window to keep the number of data samples in the window constant. If a new data sample is added, the oldest sample in the window is deleted, so that the model always learns based on the latest and most representative data, while controlling the computational complexity; The incremental dataset is preprocessed the same way as the initial training data, including outlier detection (KNN algorithm based on density peaks), missing value interpolation (Bayesian inference method), and Z-score normalization to ensure that the new data meets the model input requirements and maintains the consistency of the data distribution.

[0043] (2) Automatically trigger model updates When the prediction error exceeds the set threshold, the system automatically triggers a model update, retrains the model using incremental data, and quickly corrects the prediction results; Consider historical forecast errors, business requirements for forecast accuracy, and the model's fluctuation range to set a reasonable forecast error threshold. The following methods can be used: Statistical analysis method: Calculate the mean and standard deviation of historical prediction errors (such as MSE and RMSE), set the threshold to the mean plus several times the standard deviation (such as mean + 2 times the standard deviation). When the prediction error exceeds this value, it is considered that the model prediction performance has declined. Business experience method: Based on the actual operational requirements of wind farms, domain experts set thresholds based on their experience. For example, for short-term wind power forecasts, a model update is triggered when the RMSE exceeds 10% of the actual wind power mean.

[0044] This invention utilizes a genetic algorithm to optimize neural networks, combining efficient data preprocessing, adaptive learning, and real-time prediction techniques to create a comprehensive, highly accurate, and real-time wind power prediction system. Compared to existing technologies, this system offers significant advantages in accuracy, robustness, and real-time responsiveness, providing strong technical support for wind farm operation and scheduling, and possessing broad engineering application prospects and market potential.

[0045] Some of the data in the above formulas are dimensionless and numerically calculated. Meanwhile, the contents not described in detail in this specification belong to the prior art known to those skilled in the art.

[0046] The above embodiments are only used to illustrate the technical method of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical method of the present invention may be modified or replaced by equivalents without departing from the spirit and scope of the technical method of the present invention.

Claims

1. A wind power prediction system based on genetic algorithm optimized neural network, characterized in that: include: Data acquisition module, which obtains historical power data of wind farms and their corresponding meteorological data in real time; The data preprocessing module cleans, detects outliers, and normalizes the collected historical power data and meteorological data, eliminating abnormal data and filling missing values to generate a high-quality training data set; Genetic algorithm optimization module, which uses genetic algorithm to perform global search and optimization on the initial weights, thresholds and network topology of the neural network, uses fitness function to evaluate the network performance of each iteration, and obtains the optimal parameter combination through selection, crossover and mutation operations; The neural network training module uses the preprocessed training data to perform multiple iterative training on the neural network model optimized by the genetic algorithm, and continuously adjusts the model parameters through the validation set; The prediction and output module inputs real-time meteorological data into the trained neural network model to predict wind power and visualize the prediction results.

2. The wind power prediction system based on genetic algorithm optimized neural network according to claim 1, characterized in that: The data pre-processing module cleans the historical power data and meteorological data in the following manner: Confirm the outliers of the same type of data collected by different nodes at the same time, confirm the total number of different types of data N, and confirm the corresponding type of data x i The associated local density is: Confirm the local density P j , where χ(……) is a unit step function, when (d ij -dc) < 0, the value is 1, otherwise it is 0, d ij Represents the corresponding type data x i with x j The distance between, where dc is the cutoff distance, is obtained by: ; Use: (P j ×C1)÷(d ij × C2) = Z j Determine the comprehensive evaluation value Z j and the comprehensive evaluation value Z j Compare and check with the preset threshold: Will satisfy: Z j The values ≤Y1 are calibrated as abnormal values, otherwise no calibration is performed, where Y1 is the preset value, and C1 and C2 are preset fixed coefficient factors; For the determined x i , extract other multi-dimensional data associated with this data, and for each set of multi-dimensional data and this x i The Pearson correlation coefficient was calculated to obtain the determined value. If there was a set of determined values that satisfied the following: determined value ≥ 0.7, it was considered valid data and was retained; otherwise, it was discarded.

3. The wind power prediction system based on genetic algorithm optimized neural network according to claim 2, characterized in that: The specific method of filling missing values in the data preprocessing module is: Assume that a single data follows a multivariate normal distribution N(μ,Σ), where μ is the mean vector and Σ is the covariance matrix. The data is divided into the observed part Xo and the missing part Xm, forming a data vector = (Xo,Xm); Based on the observed data Xo, the posterior distribution of the distribution parameters μ and Σ is calculated using the Bayesian estimation method; Based on the known posterior distribution of parameters, according to the properties of multivariate normal distribution, calculate the conditional distribution P(Xm|Xo) of the missing value Xm under the given Xo condition; this conditional distribution is also a normal distribution, and its mean and covariance are derived from Xo, μ and Σ; The conditional distribution P(Xm|Xo) of the missing values is sampled to obtain multiple interpolation values. To reduce interpolation bias, the mean or median of the multiple interpolation values is selected as the final missing value interpolation result.

4. The wind power prediction system based on genetic algorithm optimized neural network according to claim 3, characterized in that: The data preprocessing module performs normalization processing in the following specific manner: Normalize all data of the same type in the data set and map them to the interval [0,1]; The original data is converted into a set of new unrelated features through linear transformation. These principal components are sorted according to the size of the variance. The larger the variance, the more information it contains. During the calculation process, the data is first centered, and then the covariance matrix is calculated. The eigenvectors and eigenvalues are obtained through eigenvalue decomposition. The first k eigenvectors with larger eigenvalues are selected to construct a projection matrix, and the original data is projected into the new feature space to achieve dimensionality reduction. After feature extraction, the processed data is divided into a training set and a validation set in a ratio of 7:

3. The training set is used for model parameter learning, allowing the model to learn the relationship between features and target variables from the data. The validation set is used to evaluate model performance during model training, adjust hyperparameters, and avoid model overfitting.

5. The wind power prediction system based on genetic algorithm optimized neural network according to claim 1, characterized in that: The genetic algorithm optimization module adopts multiple fitness functions to comprehensively evaluate the prediction error, model complexity and generalization ability to ensure global search during the optimization process.

6. The wind power prediction system based on genetic algorithm optimized neural network according to claim 1, characterized in that: The neural network training module adopts a multi-layer feedforward neural network, combined with an adaptive learning rate and early stopping mechanism to prevent overfitting, and dynamically adjusts the network structure during the model training process to improve computational efficiency.

7. The wind power prediction system based on genetic algorithm optimized neural network according to claim 1, characterized in that: The data preprocessing module performs outlier detection through the KNN algorithm based on density peaks, removes noise points in wind power data, and fills data gaps with a missing value interpolation algorithm based on Bayesian inference to ensure data integrity and accuracy.

8. The wind power prediction system based on genetic algorithm optimized neural network according to claim 1, characterized in that: The prediction and output module supports short-term and ultra-short-term wind power prediction, can dynamically adjust the prediction period according to the operation demand of the wind farm, and provide the prediction results to the wind farm scheduling system through the data interface for decision support.

9. The wind power prediction system based on genetic algorithm optimized neural network according to claim 1, characterized in that: The neural network model used in the neural network training module is a multi-layer perceptron or a long short-term memory network, and the optimal network structure is automatically selected according to the data characteristics to improve the modeling ability of time series data.

10. The wind power prediction system based on genetic algorithm optimized neural network according to claim 1, characterized in that: Also includes: The adaptive update module has the ability of online learning and adaptive updating. It can dynamically adjust the neural network parameters according to the new data collected in real time, and continuously improve the long-term stability and accuracy of the prediction model.

Images