A Physical Knowledge-Guided Intelligent Wind Power Prediction Method and System

Abstract

The present invention discloses an intelligent wind power prediction method and system guided by physical knowledge. The method first collects wind farm data and performs preprocessing, and divides the preprocessed data into a training set, a test set, and a validation set. Secondly, a wind power prediction model is constructed and physical prior knowledge is incorporated, and the data in the training set is input into the wind power prediction model for training. Finally, the trained wind power prediction model is used to perform multi-step prediction through the data in the test set and the validation set to obtain the predicted power. The system includes a data acquisition unit, a data preprocessing unit, a model construction unit, a model training unit, and a model prediction unit. The present invention improves the prediction accuracy and generalization ability in complex environments, can achieve more accurate prediction of wind power of wind farms under different meteorological conditions, and at the same time enhances the robustness of the prediction.

Description

Technical Field

[0001] The present invention relates to the technical field of wind farm wind power prediction, and particularly relates to an intelligent wind power prediction method and system guided by physical knowledge. Background Art

[0002] With the transformation of the global energy structure and the increasingly serious problem of climate change, wind energy, as a clean and renewable energy source, has received extensive attention in its development and utilization. As the main utilization method of wind energy, wind power occupies an increasingly important position in the energy production around the world. However, the output of wind power is affected by various factors such as wind speed, wind direction, and air density, and has great uncertainty and volatility. These characteristics bring certain challenges to the management of wind farms and the dispatching of power grids.

[0003] Traditional wind power prediction methods mainly include two categories: physical models and statistical models. Physical models predict wind speed and wind direction by simulating the physical processes of the atmosphere, and then predict wind power. However, such models usually have complex calculations, require a large amount of meteorological data and high computing power. Statistical models such as time series analysis and regression analysis, although they simplify the calculation process to a certain extent, often ignore the complex non-linear relationship between wind power output and environmental conditions, resulting in limited prediction accuracy.

[0004] In recent years, with the development of artificial intelligence technology, wind power prediction methods based on deep learning have been widely studied. These methods can capture the non-linear relationship between wind power and related factors by learning the patterns in historical data, improving the accuracy of prediction. However, when dealing with wind power data, existing deep learning often ignores the spatial correlation between each wind turbine in the wind farm and the physical constraints of wind power generation, which may lead to insufficient practicality and reliability of the prediction results.

[0005] Therefore, there is an urgent need to develop a new wind power prediction method that can integrate the spatial relationship inside the wind farm and the physical constraints of wind power generation to improve the accuracy and reliability of wind power prediction. Summary of the Invention

[0006] The present invention provides an intelligent wind power prediction method and system guided by physical knowledge, which can effectively simulate the interaction between each wind turbine in the wind farm and its response to environmental factors such as wind speed and wind direction, so as to achieve more accurate and stable wind power prediction.

[0007] An intelligent wind power prediction method guided by physical knowledge includes:

[0008] S1. Collect wind farm data and perform preprocessing;

[0009] S2. Divide the data preprocessed in S1 into a training set, a test set, and a validation set;

[0010] S3. Construct a wind power prediction model and incorporate physical prior knowledge. Input the data of the training set processed in S2 into the wind power prediction model for training;

[0011] S4. Use the trained optimal wind power prediction model for multi-step prediction, and obtain the predicted power through the data in the test set and the validation set.

[0012] In this embodiment, the data preprocessing includes: data filtering, feature extraction, and data standardization. Data filtering eliminates abnormal data, feature extraction selects key features, and data standardization eliminates the difference in dimension.

[0013] In this embodiment, the data division is to divide the wind power data set into a training set, a validation set, and a test set.

[0014] In this embodiment, the wind power prediction model includes:

[0015] S3.1 Build a cross-temporal convolutional neural network. The cross-temporal convolutional neural network adopts an encoder-decoder architecture, and its core is the sampling convolution and cross-learning mechanism; the cross-temporal convolutional neural network is constructed by arranging multiple multi-resolution feature extraction modules into a binary tree structure; the multi-resolution feature extraction module essentially extracts multi-level temporal features from the time series through the mechanism of downsampling-convolution-interaction, and fuses interactive learning to make up for the information loss in the downsampling process, so as to achieve more efficient time series modeling and prediction; the interactive learning is to exchange information between the downsampled subsequences, and project the input sequence and onto the hidden state by two different 1D convolutional modules and respectively, and transform it through the function, and perform interaction with through element-wise product ( and ); the multi-resolution feature extraction module downsamples the original sequence into two odd and even sequences by retaining the elements with odd numbers and the elements with even numbers in the input feature and respectively, and incorporates cross-learning. The formula is as follows:

[0016]

[0017] In the formula, is the even subsequence, is the odd subsequence, and is a scaling feature, and is an updated feature of the output, The Hadamard product is the element-wise product, is the exponential function transformation, 、 、 and are different 1D convolution modules respectively, indicates that the training adaptively learns whether to use addition or subtraction operations to optimize the performance.

[0018] S3.2 Physical prior knowledge: Two physical prior knowledge are extracted according to the physical characteristics of wind power and the relationship between wind power and wind speed;

[0019] The first physical prior knowledge is integrated into a clipping module, which is used to limit the output of the wind power prediction model within a reasonable range during training and testing; According to physical laws, in practical applications, the value of wind power should be greater than zero, so the model output should be positive to eliminate physically unreasonable predictions. The formula is as follows:

[0020]

[0021] In the formula, is the output of the cross-space-time convolutional neural network, is the predicted value;

[0022] The second physical prior knowledge is integrated into a penalty loss module. Theoretically, the wind power generation within a certain period should fall within a certain range. According to the wind power conversion relationship between wind power and wind speed, the upper and lower bound constraint functions are deduced, and the penalty loss module is constructed by the upper and lower bound functions. The formula is as follows:

[0023]

[0024]

[0025] In the formula, is the wind speed, is the cut-in wind speed, is the rated wind speed, is the cut-out wind speed, is the model output, is the penalty loss, is the total number of samples, and other constant parameters 、 、 、 、 can be obtained by fitting with software fitting tools.

[0026] S3.3 Model Training: The loss function of the wind power prediction model is composed of the loss term of the cross - spatio - temporal convolutional neural network and the penalty loss term that integrates physical prior knowledge. The formula is as follows:

[0027]

[0028]

[0029]

[0030] In the formula, is the loss function of the cross - spatio - temporal convolutional neural network wind power prediction model, is the penalty loss, is the total number of samples, is the loss function of the wind power prediction model that integrates physical prior knowledge, is the true value, and are the predicted value and the model output respectively.

[0031] The multi - step prediction is to perform power prediction using recursive multi - step prediction. The recursive multi - step prediction first initializes the multi - step prediction. The result of the first - step prediction is concatenated with the original data along the time dimension to obtain the initial input for the next - step prediction. Then, multi - step recursive prediction is carried out. When the number of prediction steps reaches the set value, the prediction stops. The input of the multi - step recursive prediction is a sequence with a time window length of T1, and the output is the predicted value for the next T2 steps.

[0032] An intelligent wind power prediction system guided by physical knowledge includes:

[0033] Data Acquisition Unit: The data acquisition module uses sensors to collect wind farm data;

[0034] Data Pre - processing Unit: The data pre - processing module is used to pre - process the wind farm data to obtain high - quality model input data;

[0035] Model Construction Unit: The model construction module is used to construct a wind power prediction model and integrate physical prior knowledge at the same time;

[0036] Model Training Unit: The model training module is used to construct the loss function of the wind power prediction model and train the wind power prediction model through the wind farm data and the loss function to obtain a trained prediction model;

[0037] Model prediction unit: The model prediction module is used to input the wind farm data for measurement into the trained wind power prediction model to obtain the future multi-step wind power prediction for wind power prediction.

[0038] Advantages of the present invention:

[0039] As can be seen from the above technical solutions, the present invention provides an intelligent wind power prediction method and system guided by physical knowledge. This method constructs a wind power prediction model that can simultaneously consider the physical characteristics of the wind farm and the characteristics of historical data by combining physical prior knowledge in the wind power field and deep learning. The wind power prediction model ensures the accuracy of model prediction by introducing physical prior knowledge in the wind power field. In addition, the wind power prediction model effectively extracts and fuses the mutual influences and spatio-temporal characteristics among multiple wind turbines by using the downsampling-convolution-interaction mechanism, greatly improving the prediction accuracy and generalization ability of the model in complex environments. Based on the work done in this research, it is possible to achieve more accurate prediction of wind power in wind farms under different meteorological conditions and at the same time improve the robustness of the prediction. Description of the drawings

[0040] Figure 1 It is a flowchart of an intelligent wind power prediction method guided by physical knowledge provided by an embodiment of the present invention;

[0041] Figure 2 It is a schematic diagram of the model logic of an intelligent wind power prediction method guided by physical knowledge provided by an embodiment of the present invention;

[0042] Figure 3 It is a schematic diagram of the prediction effect of an intelligent wind power prediction method guided by physical knowledge provided by an embodiment of the present invention;

[0043] Figure 4 It is a schematic diagram of the structure of an intelligent wind power prediction system guided by physical knowledge provided by an embodiment of the present invention. Detailed implementation manners

[0044] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

[0045] The embodiments of the present invention disclose an intelligent wind power prediction method and system guided by physical knowledge, aiming to achieve more accurate prediction of wind power in wind farms under different meteorological conditions.

[0046] Please refer to Figure 1 , this application proposes an intelligent wind power prediction method guided by physical knowledge, including:

[0047] Step 101: Collect wind farm data and perform preprocessing;

[0048] Specifically, in this embodiment, the data preprocessing in step 101 includes:

[0049] Data filtering: According to the normal operating parameters of the wind turbine, eliminate the data with abnormal wind speed to ensure that the model training is carried out on valid data;

[0050] Feature extraction: Select wind speed, wind direction, and historical power as the input features of the wind power prediction model;

[0051] Data standardization: In order to eliminate the dimensional difference between different features, the data is standardized. The formula is as follows:

[0052]

[0053] In the formula, X is the original data, μ is the mean value, σ is the standard deviation, is the standardized data.

[0054] Step 102: Divide the data preprocessed in step 101 into a training set, a test set, and a validation set;

[0055] Specifically, in this embodiment, the data is divided by proportionally dividing the wind farm dataset into a training set (70%), a test set (20%), and a validation set (10%), training set (70%). The training set contains a large amount of historical data. The model adjusts its own parameters by learning the patterns and relationships in these data to achieve the best prediction performance. The validation set is used to evaluate and adjust the performance of the model during the training process, which helps to optimize the structure and training strategy of the model and ensure that the model has good generalization ability on unseen data. The test set is used to evaluate the final performance of the model after the model training is completed. The test set is the data that the model has not come into contact with during the training process. By evaluating on the test set, the actual prediction ability of the model can be objectively measured. Use the sliding window method to input the data into the wind power prediction model for training.

[0056] Step 103: Build a wind power prediction model and incorporate physical prior knowledge, and input the data processed in step 102 into the wind power prediction model for training;

[0057] Step 104: Use the trained optimal wind power prediction model to perform multi-step prediction to obtain the predicted power.

[0058] Please refer to Figure 2 , in this embodiment, the wind power prediction model used in step 103 of the present application includes:

[0059] Build a cross - spatio - temporal convolutional neural network. The cross - spatio - temporal convolutional neural network adopts an encoder - decoder architecture, and its core is sampling convolution and cross - learning mechanism; the cross - spatio - temporal convolutional neural network is constructed by arranging multiple multi - resolution feature extraction modules into a binary tree structure; the multi - resolution feature extraction module essentially extracts multi - level temporal features from the time series through a mechanism of downsampling - convolution - interaction, and fuses interactive learning to make up for the information loss in the downsampling process, so as to achieve more efficient time series modeling and prediction; the interactive learning is to exchange information between the downsampled subsequences, and and are respectively projected into the hidden state by two different 1D convolutional modules and , and are transformed through the function, and interact with and and through element - wise product ( ); the multi - resolution feature extraction module downsamples the original sequence into two odd - even sequences and by respectively retaining the elements with odd numbers and the elements with even numbers in the input feature , and incorporates cross - learning. The formula is as follows:

[0060]

[0061] In the formula, is the even subsequence, is the odd subsequence, and are the scaled features, and are the updated features of the output, Hadamard product is the element - wise product, is the exponential function transformation, , , and are different 1D convolutional modules respectively, represents that the training adaptively learns whether to use addition or subtraction operations to optimize the performance.

[0062] In this embodiment, the function of the cross - spatio - temporal convolutional neural network is to divide a sequence into an odd sequence and an even sequence after interaction. From the perspective of sampling frequency, it is similar to performing two regular down - sampling operations; the cross - spatio - temporal convolutional neural network adopts a tree - like structure to nest the multi - resolution feature extraction modules. There are 2L−1 multi - resolution feature extraction modules at the L - th layer. After sampling through L layers, a total of 2L subsequences will be obtained; the lengths of these sequences will only be 1 / 2 of the original sequence length. L , after using concatenation and alignment splicing operations, the total sequence length is equal to the original sequence length.

[0063] In this embodiment, a significant advantage of the cross - spatio - temporal convolutional neural network is that each multi - resolution feature extraction module has local and global views of the entire time series, facilitating the extraction of useful time features; after all the down - sampling - convolution - interaction operations, the extracted features are rearranged into a new sequence representation and added to the original time series, and a fully - connected network is used as the decoder for prediction.

[0064] Physical prior knowledge: Two physical prior knowledge are extracted based on the physical characteristics of wind power and the relationship between wind power and wind speed:

[0065] The first physical prior knowledge is integrated into a clipping module, which is used to limit the output of the wind power prediction model within a reasonable range during training and testing; according to physical laws, in practical applications, the value of wind power should be greater than zero. Therefore, the model output should be positive to eliminate physically unreasonable predictions. The formula is as follows:

[0066]

[0067] In the formula, is the output of the cross - spatio - temporal convolutional neural network, is the predicted value.

[0068] The second physical prior knowledge is integrated into a penalty loss module. Theoretically, the wind power generation within a certain period should fall within a certain range. According to the wind power conversion relationship between wind power and wind speed, an upper and lower bound constraint function is derived, and a penalty loss module is constructed from the upper and lower bound functions. The formula is as follows:

[0069]

[0070]

[0071] In the formula, is the wind speed, is the cut - in wind speed, is the rated wind speed, is the cut - out wind speed, is the model output. is the penalty loss, is the total number of samples, and other constant parameters can be obtained by fitting with a software fitting tool.

[0072] Model training: The loss function of the wind power prediction model is composed of the loss term of the cross - space - time convolutional neural network and the penalty loss term that fuses physical prior knowledge, and the formula is as follows:

[0073]

[0074]

[0075]

[0076] In the formula, is the loss function of the cross - space - time convolutional neural network wind power prediction model, is the penalty loss, is the total number of samples, is the loss function of the wind power prediction model that fuses physical prior knowledge, is the true value, and are the predicted value and the model output respectively.

[0077] The wind power prediction model in this embodiment uses the training set data to train the model, adjusts the model parameters through an optimization algorithm (Adam optimizer) to minimize the loss function on the training set. After training is completed, the validation set is used to evaluate the performance of the model. The hyperparameters of the model are adjusted according to the performance on the validation set to prevent overfitting. After the model training and validation are completed, the test set is used to evaluate the final performance of the model. The results of the test set are used to measure the prediction ability of the model in practical applications. The input of the model is data of 16 time steps, and the output is the state data of the 17th step. The number of hidden layers of the wind power prediction model is 5, the size of the convolutional kernel is 3, the learning rate is 0.01, and the batch size is 32.

[0078] The wind power prediction model in this embodiment uses an early stopping mechanism to prevent the model from overfitting and achieve the best performance on the validation set. The core idea of the early stopping mechanism is that if the performance of the model on the validation set does not improve significantly after a certain number of training epochs, the training is stopped in advance. The patience value of the early stopping mechanism is set to 10. The schematic diagram of the prediction effect is as Figure 3 shown.

[0079] In this embodiment, the MAE and R 2 of the original wind power prediction model are 178.604 and 0.957 respectively; the MAE and R of the wind power prediction model after fusing physical prior knowledge2 They are 157.541 and 0.9611 respectively. Compared with the original model, the MAE is reduced by about 11.8%, and the R 2 is increased by about 0.43%, showing a significant improvement compared with the original wind power prediction model. The results show that the wind power prediction model integrating physical prior knowledge can more deeply reveal the internal variation law of the wind power sequence and improve the overall prediction accuracy.

[0080] In this embodiment, the model prediction in step 104 of the present application includes:

[0081] The multi-step prediction is to use recursive multi-step prediction for power prediction. The recursive multi-step prediction first initializes the multi-step prediction, splices the first-step prediction result and the original data along the time dimension to obtain the initial input for the next-step prediction. Then, multi-step recursive prediction is performed. When the number of prediction steps reaches the set value, the prediction stops. The input of the multi-step recursive prediction is a sequence with a time window length of T1, and the output is the predicted value for the next T2 steps. The number of recursive times of the multi-step recursive prediction is pred_step - 1 times, and pred_step = 16.

[0082] Please refer to Figure 4 , in the embodiment of the present invention, an intelligent wind power prediction system based on physical knowledge guidance is provided, including:

[0083] Data acquisition unit 401: The data acquisition module uses sensors to collect wind farm data;

[0084] Data preprocessing unit 402: The data preprocessing module is used to preprocess the wind farm data to obtain high-quality model input data;

[0085] Model construction unit 403: The model construction module is used to construct a wind power prediction model and integrate physical prior knowledge at the same time;

[0086] Model training unit 404: The model training module is used to construct the loss function of the wind power prediction model, and train the wind power prediction model through the wind farm data and the loss function to obtain a trained prediction model;

[0087] Model prediction unit 405: The model prediction module is used to input the measured wind farm data into the trained wind power prediction model to obtain the future multi-step wind power prediction for wind power prediction.

[0088] The present invention constructs an intelligent wind power prediction method guided by physical knowledge, effectively integrating the local information of wind turbines in a wind farm and global environmental factors, thereby achieving high accuracy and efficiency in wind power prediction. Specifically, the wind power prediction model that incorporates physical prior knowledge can explore the physical characteristics of each wind turbine or wind farm site itself, such as local influencing factors like wind speed, wind direction, and air density, while analyzing and processing the interactions between wind turbines and their responses to the surrounding environment; this model can not only more comprehensively reflect the complex correlations and mutual influences among various sites in the region, but also fully consider the influence of global environmental factors while retaining the local characteristics of each wind turbine, thus significantly improving the prediction accuracy of wind power.

[0089] Within the field of the present invention, those skilled in the art should understand that, according to the various components and algorithm steps shown in the disclosed embodiments of the present invention, they can be implemented by electronic hardware, computer software, or a combination thereof. To clearly demonstrate the interchangeability of hardware and software, these components and steps have been described in terms of their general functions. The implementation of these functions, whether through hardware or software, depends on the specific technical application environment and design constraints. Skilled technicians can choose different implementation means according to each specific application environment to achieve these functions, and such a choice should not be regarded as departing from the scope of the present invention.

Claims

1. A wind power intelligent prediction method based on physical knowledge, characterized in that: The following steps are involved: S1. Collect wind farm data and perform preprocessing; S2. Divide the preprocessed data into training set, test set and validation set; S3. Construct a wind power prediction model, integrate physical prior knowledge, and input the data of the training set into the wind power prediction model for training; The specific implementation process of the wind power prediction model is as follows: S3.1 builds a cross-spatiotemporal convolutional neural network, extracts data features of the training set and calculates the loss; S3.2 Extract two physical prior knowledge based on the physical characteristics of wind power and the relationship between wind power and wind speed; the two physical prior knowledge are as follows: The first physical prior knowledge is integrated into the clipping module. According to the laws of physics, in practical applications, the value of wind power should be greater than zero, so the output should be positive to eliminate physically unreasonable predictions. The formula is as follows: In the formula, is the output of the cross-spatiotemporal convolutional neural network, is the predicted value; The second physical prior knowledge is integrated into a penalty loss module. According to the wind power conversion relationship between wind power and wind speed, the upper and lower limit constraint functions are introduced. The penalty loss module is constructed by the upper and lower limit functions. The formula is as follows: In the formula, v i is the wind speed, v in is the cut-in wind speed, v rat is the rated wind speed, v out is the cut-out wind speed, is the model output, MSE plt is the penalty loss, N is the total number of samples, and other constant parameters a n 、a' n , b n , b n ',ω is obtained by fitting with the software fitting tool; S3.3 construct the loss function of the wind power prediction model and perform training; S4. Use the trained wind power prediction model to perform multi-step predictions using the data in the test set and validation set to obtain the predicted power.

2. The method for intelligent wind power prediction based on physical knowledge guidance according to claim 1 is characterized in that: The cross-spatiotemporal convolutional neural network is an encoder-decoder architecture, including sampling convolution and cross-learning mechanisms; The cross-spatiotemporal convolutional neural network is constructed by arranging multiple multi-resolution feature extraction modules into a binary tree structure; the essence of the multi-resolution feature extraction module is to extract multi-level time features from the time series through the mechanism of downsampling-convolution-interaction, and integrate interactive learning to compensate for the information loss in the downsampling process; The interactive learning is to exchange information between the downsampled subsequences. even and F odd Two different 1D convolution modules φ and ψ are used to project the hidden state, transform it through the exp function, and then multiply it with F by element-wise product ⊙ even and F odd The multi-resolution feature extraction module downsamples the original sequence into two odd-even sequences F by retaining the elements with odd numbers and the elements with even numbers in the input feature F respectively. odd and F even , and integrate cross-learning, the formula is as follows: In the formula, F even is an even subsequence, F odd is an odd subsequence, and is the scaling feature, F' even and F' odd is the updated feature of the output, exp is the exponential function transformation, φ, ψ, ρ and η are different 1D convolution modules, and ± indicates whether the training adaptive learning uses addition or subtraction operations.

3. The method for intelligent wind power prediction based on physical knowledge guidance according to claim 2 is characterized in that: The loss function of the wind power prediction model is the MSE of the cross-spatiotemporal convolutional neural network. data Loss term and MSE integrating physical prior knowledge plt The penalty loss term is composed of the following formula: MSE=MSE data +MSE plt In the formula, MSE data is the loss function of the cross-spatial convolutional neural network wind power prediction model, MSE plt is the penalty loss, N is the total number of samples, MSE is the loss function of the wind power prediction model integrating physical prior knowledge, y i is the true value.

4. The method for intelligent wind power prediction based on physical knowledge guidance according to claim 3 is characterized in that: The multi-step prediction is to use recursive multi-step prediction to perform power prediction. The recursive multi-step prediction first performs multi-step prediction initialization, splices the first step prediction result with the original data along the time dimension, and obtains the initial input for the next step prediction; then multi-step recursive prediction is performed, and then multi-step recursive prediction is performed. When the number of prediction steps reaches a set value, the prediction is stopped; the input of the multi-step recursive prediction is a sequence with a time window length of T1, and the output is the prediction value of the future T2 steps.

5. A wind power intelligent prediction system guided by physical knowledge, used to implement the wind power intelligent prediction method according to any one of claims 1 to 4, characterized in that: Includes the following modules: Data acquisition unit: uses sensors to collect wind farm data; Data preprocessing unit: used for preprocessing wind farm data; Model building unit: used to build a wind power prediction model while incorporating physical prior knowledge; Model training unit: used for constructing the loss function of the wind power prediction model, training the wind power prediction model through the wind farm data and the loss function, and obtaining a trained prediction model; Model prediction unit: used to input the measured wind farm data into the trained wind power prediction model, obtain the future multi-step wind power prediction, and perform wind power prediction.

Images