A long-distance wind power generation power prediction method based on a two-way timeline

Abstract

The present invention discloses a long-distance wind power generation power prediction method based on a two-way timeline, which specifically includes: First, perform data preprocessing operations such as data normalization, data cleaning, data supplementation, and data screening on the meteorological data, equipment monitoring data, and wind turbine basic data of the wind farm. Sort the processed data by time and send it into the calculation model. Then construct a two-way Transformer model, and through high-dimensional feature calculation and self-attention mechanism, perform weight ratio matching and feature extraction on the input features, so that part of the data can be abstracted by the model into important features for output power prediction. The present invention can not only simulate the unidirectional influence of meteorological changes on wind turbines, but also reverse the reduction of errors through the target output power at a determined future time point, and finally realizes a more accurate prediction of the output power of wind power generation, and improves the accuracy of long-distance wind power generation power prediction without reducing the prediction efficiency.

Description

Technical Field

[0001] The present invention belongs to the technical field of wind power generation, and particularly relates to a long-distance wind power generation power prediction method based on a two-way timeline. Background Art

[0002] With the continuous development of the new energy industry, the total installed capacity and installed capacity of wind power in China have increased rapidly. The intelligentization of wind turbines will be an important development trend in the wind power industry. On the one hand, the intelligent development of wind turbines requires the traditional manufacturing process to pass as the basis, and on the other hand, it needs to fully integrate Internet technology innovation and apply new generation information technologies such as big data and sequence prediction to finally achieve the background monitoring and full life cycle management of remote wind turbines, ensuring the operation status of smart wind farms. The intelligent power prediction and monitoring of wind power generation, as the "power generation heart" of a digital smart wind farm, plays a crucial role in the subsequent dispatching and grid connection of the output power of the wind farm. Therefore, how to efficiently and accurately predict the power generation power of wind power generation has become a key concern in the wind power field.

[0003] As a new generation information technology, big data analysis technology can combine real-time data stream analysis with historical related data, and help the system predict and prevent possible interruptions and performance problems in future operations by modeling data scenarios.

[0004] At present, the wind power generation power prediction methods mainly focus on statistical models and deep neural network models, such as Support Vector Machines (SVM), Long Short-Term Memory (LSTM), Temporal Convolutional Network (TCN), etc., and certain prediction effects have been achieved. However, in the current model training and use processes, a method relying on timeline modeling is adopted. Although this unidirectional sequence modeling algorithm can simulate the influence of meteorological changes on the power generation power of a wind farm, the error at each step will accumulate along the timeline. Especially during the model prediction process, this error will accumulate to the maximum at the final prediction moment, seriously affecting the subsequent power generation grid connection dispatching of the wind farm. Summary of the Invention

[0005] To solve the above problems, the present invention provides a long-distance wind power generation power prediction method based on a two-way timeline. By using the two-way timeline modeling method and relying on a "[MASK] cloze" - like training process, while ensuring the task consistency during model training and prediction processes, more accurate wind power generation power prediction is achieved, effectively improving the defects of the existing unidirectional power prediction algorithm and assisting the wind farm to make more effective decisions on subsequent power grid connection dispatching.

[0006] The Transformer is a model that uses the attention mechanism to improve the model training speed. It uses an encoder-decoder architecture and is a deep learning model completely based on the self-attention mechanism. Because it is suitable for parallel computing and the complexity of its own model, it is higher than the previously popular Recurrent Neural Network (RNN) in terms of accuracy and performance. Based on this, the present invention proposes a long-distance wind power prediction method based on a bidirectional timeline Transformer architecture. Its network model structure includes an input layer, a bidirectional Transformer layer, a reverse error smoothing layer, and a fully connected layer. The algorithm of the present invention can effectively solve the problem of large cumulative errors in the existing unidirectional timeline model, and effectively improve the prediction accuracy of long-distance output power while not significantly reducing the model calculation time.

[0007] To achieve the above object, the technical solution of the present application is as follows:

[0008] A long-distance wind power prediction method based on a bidirectional timeline, specifically including:

[0009] Step 1: Perform data preprocessing operations such as data normalization, data cleaning, data supplementation, and data screening on the wind farm meteorological data, equipment monitoring data, and wind turbine basic data. Sort the processed data according to time and send it into the bidirectional Transformer model;

[0010] Step 2: Send the normalized features into the bidirectional Transformer model, transform the data into high-dimensional feature vectors, and then perform feature calculation and attention weight allocation;

[0011] Step 3: Calculate the output power prediction result of the last time point through the bidirectional Transformer model, and use this result as the standard output. The reverse Transformer layer performs autoregressive modeling training on the entire sequence by replacing the sequence content with [MASK], which can effectively smooth the errors of the predictions at intermediate time nodes and define the prediction interval;

[0012] Step 4: Through the fully connected layer, project the hidden layer features onto the conventional output space of power prediction to obtain the wind power prediction results at each time point.

[0013] Further, the specific implementation method of the said Step 2 is as follows:

[0014] Step 21: Construct a bidirectional Transformer model, perform weight ratio matching and feature extraction on the input features through high-dimensional feature calculation and the self-attention mechanism, and realize that some data is abstracted by the model as important features for output power prediction.

[0015] Technical effects of the present invention:

[0016] During the training process of the present invention, a two-way timeline is adopted to construct the hidden connection between meteorological features and power generation power. This method can not only simulate the one-way influence of meteorological changes on wind power generation, but also reverse the reduction of errors through the target output power at a determined future time point. During the model training process, the reverse Transformer layer changes the training target of the model from predicting the power generation power at the next moment to predicting the power generation power at the [MASK] position by replacing a part of the power generation power in the input sequence with [MASK]. This method can more effectively model the autocorrelation of the overall power sequence, improve the long-distance continuity fitting ability of the model for the power prediction sequence, and finally achieve more accurate prediction of wind power generation output power. This method can improve the prediction accuracy of long-distance wind power generation power without reducing the prediction efficiency.

[0017] By using the long-distance wind power generation power prediction method based on the two-way timeline Transformer architecture, the problem of large cumulative errors in the existing one-way timeline model can be effectively solved, and while not significantly reducing the model calculation time, the prediction accuracy of the output power can be effectively improved, assisting the wind farm to make more effective subsequent power grid connection scheduling and command decisions. Description of the drawings

[0018] Figure 1 It is a schematic flow diagram of a long-distance wind power generation power prediction method based on a two-way timeline. Specific implementation manners

[0019] The present invention will be further described in detail below in conjunction with the drawings and specific embodiments, and this is taken as an example to further describe the present application.

[0020] Embodiment 1:

[0021] In the process of wind power generation prediction in a wind farm, a wind power generation prediction method based on a bidirectional timeline Transformer architecture is used to perform high-dimensional feature mapping on multiple features in the input source, and to weight the importance of power prediction features. Through bidirectional modeling, the problem of error accumulation over time in the power prediction process is avoided. Referring to Figure 1, the present application provides a long-distance wind power generation prediction method: first, data preprocessing operations such as data normalization, data cleaning, data supplementation and data screening are performed on the meteorological data of the wind farm, equipment monitoring data, and basic data of wind turbines. The processed data is sorted by time and sent to the calculation model. Then a bidirectional Transformer model is constructed, and the input features are weighted and extracted through high-dimensional feature calculation and self-attention mechanism, so that some data is abstracted as important features for output power prediction with the help of the model. Then, the output power prediction result at the last moment is calculated through the bidirectional Transformer model, and the result is used as the standard output. The prediction value of the intermediate time node is reversely calculated, the error is smoothed, and the prediction interval is defined. Finally, through the fully connected layer, the features of the hidden layer are projected onto the conventional output space of power prediction to obtain the wind power prediction results at each time point. This application uses a modeling method based on a two-way timeline to predict wind power. The application method can not only simulate the unidirectional impact of meteorological changes on wind power prediction, but also reduce errors in reverse by determining the target output power at a future time point, and finally achieve a more accurate prediction of wind power output, and improve the accuracy of long-distance wind power prediction without reducing the prediction efficiency.

[0022] Embodiment 2:

[0023] The present invention is described in detail below in conjunction with embodiments and drawings so that those skilled in the art can implement the invention according to the description.

[0024] This example uses Pycharm as the development platform, Python as the development language, and Pytorch as the underlying development architecture. The following is the specific process:

[0025] Step 1: Perform data preprocessing operations such as data normalization, data cleaning, data supplementation and data screening on the wind farm meteorological data, equipment monitoring data and wind turbine basic data. Sort the processed data by time and send them to the calculation model.

[0026] Step 11: Standardize and clean meteorological data, such as wind speed, temperature, humidity, air pressure, etc., and remove abnormal data; normalize wind direction, rudder angle and other features, and transform their mapping space into a simple constant space using sin, cos and other functions;

[0027] Step 12: Sort the preprocessed data according to the timeline and then feed it into the model batch by batch.

[0028] Step 2: Construct a bidirectional Transformer model to perform feature dimension elevation on the input features and calculate the hidden relationship between each feature and the output power using the high-dimensional space.

[0029] Step 21: Elevate the dimension of the input features through vector dot product and calculate the hidden relationship between the features and the final output power through the self-attention layer;

[0030] Step 22: Update the feature weights through backpropagation.

[0031] Step 3: Calculate the predicted result of the output power at the last time point through the bidirectional Transformer model, use this result as the standard output, and perform error smoothing and prediction interval delimitation on the intermediate time node predictions in reverse.

[0032] Step 31: Perform "cloze test" style input feature processing, that is, randomly cover the historical power generation power with [mask] in the input features. The coverage probability is 15% for each input feature;

[0033] Step 32: Give priority to the operation of replacing the original content of the [mask] sequence for the time point features at the 3 / 4 position in chronological order. This not only conforms to the natural property that the features of the subsequent time points are more uncertain in accordance with the timeline order, but also can fuse the relatively complete feature sequence in the front part and the features of the last small part of time points for bidirectional training to smooth the errors in the intermediate time node predictions;

[0034] Step 33: Use the bidirectional Transformer architecture to model the input, perform self-correlation learning on the sequence, predict the original value of the masked part, and repeat this training process.

[0035] Step 4: Project the features of the hidden layer onto the conventional output space of power prediction through the fully connected layer to obtain the predicted result of the wind power generation power at each time point.

[0036] Step 41: Determine the predicted value of the wind power generation power at future time points and the length of the predicted time points by limiting the output dimension of the fully connected layer.

[0037] The above is only a preferred specific implementation manner of the present invention, but the protection scope of the present invention is not limited thereto. Any person skilled in the art within the technical scope disclosed by the present invention, according to the technical solution and inventive concept of the present invention, makes equivalent substitutions or changes, and should be covered within the protection scope of the present invention.

Claims

1. A long-distance wind power prediction method based on a two-way timeline, It is characterized in that Specifically include: Step 1: Perform data preprocessing on the wind farm meteorological data, equipment monitoring data, and wind turbine basic data; sort the processed data by time and feed them into the bidirectional Transformer model; Step 2: Send the normalized features into the bidirectional Transformer model, convert the data into a high-dimensional feature vector, and then perform feature calculation and attention weight allocation; Step 3: The output power prediction result at the last moment is calculated through the bidirectional Transformer model and used as the standard output. The reverse Transformer layer replaces the sequence content with [MASK] to perform autoregressive modeling training on the entire sequence, which can effectively smooth the error of the prediction of the intermediate time nodes and define the prediction interval; Step 4: Project the features of the hidden layer onto the conventional output space of power prediction through the fully connected layer to obtain the power prediction results at each time point; The specific implementation of step 3 is as follows: Step 31: Perform "cloze" input feature processing, that is, randomly cover the historical power generation with [mask] in the input feature; the coverage probability is 15% of each input feature; Step 32: [mask] is preferentially assigned to the time point feature that is at the end of the input sequence, and subsequent training is performed using the complete feature sequences at both ends; specifically, [mask] is preferentially replaced with the original content of the sequence, and the time point feature that is at the 3 / 4 position in the time sequence is performed, and subsequent training is performed using the complete feature sequences at both ends; Step 33: Use a bidirectional Transformer architecture to model the input, predict the original value of the masked part, and repeat the process training.

2. According to the method for predicting long-distance wind power generation based on a two-way timeline as claimed in claim 1, It is characterized in that The specific implementation of step 1 is as follows: Step 11: Standardize and clean the meteorological data, including wind speed, temperature, humidity, and air pressure characteristics, and remove abnormal data; normalize the wind direction and rudder angle characteristics, and transform their mapping space into a simple constant space using sin and cos functions; Step 12: Sort the preprocessed data by timeline and then feed it into the model in batches.

3. A long-distance wind power generation prediction method based on a bidirectional timeline according to claim 1 or 2, It is characterized in that The specific implementation of step 2 is as follows: Step 21: Upscale the input features by vector dot multiplication, and calculate the hidden relationship between the features and the final output power through the self-attention layer; Step 22: Update feature weights via backward chain derivation.

4. A long-distance wind power generation prediction method based on a bidirectional timeline according to claim 1 or 2, It is characterized in that The specific implementation of step 4 is as follows: Step 41: The input features are dimensionally upgraded by vector dot multiplication, and the hidden relationship between the features and the final output power is calculated by the self-attention layer; Step 42: Update the feature weights by backchaining differentiation.

Images