Method for constructing wind farm power prediction model based on unmanned aerial vehicle and prediction method

Abstract

The application relates to the technical field of new energy, and discloses a wind farm power prediction model construction method and prediction method based on a UAV, which comprises the following steps: obtaining historical data sets corresponding to historical icing events of a target wind farm station; inputting historical image sequences into a double-branch feature extraction model to obtain historical icing thickness sequences and historical icing shape sequences; performing feature reconstruction on the features of the historical icing thickness sequences and the historical icing shape sequences to obtain icing state fusion features; extracting time sequence dependency relationships of the historical icing state fusion feature sequences and historical power sequences to obtain time sequence change features of the icing state fusion features related to the historical power sequences; constructing a training data set based on the time sequence change features and meteorological data; inputting the training data set into a preset power prediction model which has been constructed in advance to perform model training, and obtaining a wind turbine power prediction model; the model can adapt to the accompanying relationship between icing state changes and output power changes.

Description

Technical Field

[0001] This invention relates to the field of new energy technology, specifically to a method for constructing and predicting wind farm power prediction models based on unmanned aerial vehicles (UAVs). Background Technology

[0002] Wind power generation has a high resource utilization rate and mature technology. As technology develops, costs are also decreasing, so the capacity of pile foundations has increased significantly.

[0003] Wind farms are often located in remote mountainous or forested areas with harsh weather conditions and weak communication signals. As a result, they are prone to icing in extremely cold weather, which is difficult to detect and report in a timely manner. This leads to a significant drop in the accuracy of wind power prediction for a period of time, causing great losses.

[0004] For power prediction under icing conditions, one approach is to construct a dataset corresponding to the power output of the wind turbine under icing conditions, such as temperature, humidity, and wind speed, and then train a model accordingly. Another approach is to construct a dataset corresponding to the power output of the wind turbine under icing conditions, and then train a model accordingly. Both of these methods typically map meteorological conditions or icing thickness to power output one-to-one, without considering the cumulative effect of icing on power prediction over time, resulting in inaccurate power predictions under icing conditions. Summary of the Invention

[0005] In view of this, the present invention provides a method for constructing a wind farm power prediction model and a prediction method based on UAVs, in order to solve the problem that related technologies do not consider the impact of icing conditions on power prediction over time, resulting in inaccurate predicted power under icing conditions.

[0006] In a first aspect, the present invention provides a method for constructing a wind farm power prediction model based on unmanned aerial vehicles (UAVs). The method includes: acquiring historical datasets corresponding to multiple historical icing events at a target wind farm; wherein the historical datasets include multiple historical image sequences and historical power sequences collected by UAVs at preset frequencies; the historical image sequences include multiple historical image data representing the icing state of all wind turbines installed within the target wind farm; inputting the historical image sequences into a dual-branch feature extraction model to obtain historical icing thickness sequences and historical icing shape sequences, wherein the dual-branch feature extraction model includes a first edge detection model for extracting icing thickness features and a second edge detection model for extracting icing shape features; reconstructing the features of the historical icing thickness sequences and historical icing shape sequences to obtain icing state fusion features; extracting the temporal dependency between the historical icing state fusion feature sequence and the historical power sequence to obtain temporal variation features of the icing state fusion features related to the historical power sequence; constructing a training dataset based on the temporal variation features and meteorological data; inputting the training dataset into a pre-constructed preset power prediction model for model training to obtain a wind turbine power prediction model.

[0007] As an exemplary embodiment, the wind turbine power prediction model includes at least two wind turbine power prediction sub-models, which are used for power prediction at different icing change rates. The method for constructing a wind farm power prediction model based on UAVs further includes: dividing the historical icing state fusion feature sequence according to a preset time period to obtain an icing state fusion feature sub-sequence; clustering the icing state fusion feature sub-sequence according to the icing change rate to obtain at least two historical clusters; dividing the training dataset according to the historical clusters to obtain at least two training data subsets corresponding to different icing change rates; and training different wind turbine power prediction sub-models using the training data subsets to obtain wind turbine power prediction sub-models corresponding to different icing change rates, so as to predict the power of the input data corresponding to different icing change rates.

[0008] As an exemplary embodiment, the icing change rate includes the icing thickness change rate and the icing shape change rate.

[0009] As an exemplary embodiment, the wind turbine power prediction model further includes a first correction sub-model. The first correction sub-model is used to output the first correction parameter of the short-term power prediction result. The method for constructing a wind farm power prediction model based on UAV includes: extracting each historical icing thickness value and the historical power data corresponding to the time sequence of the historical icing thickness values ​​from the historical icing thickness sequence as a first training set; inputting the first training set into the first correction sub-model for model training, and learning the correspondence between historical icing thickness values, historical power data and the first correction parameter during the training process.

[0010] As an exemplary embodiment, the wind turbine power prediction model further includes a second correction sub-model. The second correction sub-model is used to output the second correction parameter of the short-term power prediction result. The method for constructing the wind farm power prediction model based on UAV includes: extracting each historical icing static shape feature and the historical power data corresponding to the time sequence of the historical icing static shape features from the historical icing shape sequence as a second training set; inputting the second training set into the second correction sub-model for model training, and learning the correspondence between the historical icing static shape features, historical power data and the second correction parameter during the training process.

[0011] Secondly, a wind farm power prediction method based on unmanned aerial vehicles (UAVs) is provided. The method includes: acquiring a sequence of measured images and predicted meteorological data of the target wind farm, wherein the measured image sequence is acquired by the UAV at a preset acquisition frequency; inputting the measured image sequence into a dual-branch feature extraction model to obtain a sequence of measured icing thickness and a sequence of measured icing shape, wherein the dual-branch feature extraction model includes a first edge detection model for extracting icing thickness features and a second edge detection model for extracting icing shape features; reconstructing the features of the measured icing thickness and icing shape sequences to obtain measured icing state fusion features; extracting the temporal dependency of the measured icing state fusion feature sequence to obtain the actual temporal variation features of the measured icing state fusion features; and inputting the actual temporal variation features and measured meteorological data into a pre-trained wind turbine power prediction model for power prediction to obtain the wind farm power prediction result.

[0012] As an exemplary embodiment, the wind turbine power prediction model includes at least two wind turbine power prediction sub-models, which are used for power prediction at different icing thickness change rates. The wind farm power prediction method based on UAVs further includes: dividing the measured icing state fusion feature sequence according to a preset duration to obtain the measured icing state fusion feature sub-sequence; clustering the measured icing state fusion feature sub-sequence and historical clusters according to the actual icing change rate to obtain the clustering result; selecting the wind turbine power prediction sub-model corresponding to the historical cluster with a clustering degree greater than the preset clustering degree based on the clustering result, and inputting the measured icing state fusion feature sub-sequence into the wind turbine power prediction sub-model to obtain the corresponding wind turbine power prediction sub-result.

[0013] As an exemplary embodiment, the actual icing change rate includes the icing thickness change rate and the icing shape change rate.

[0014] As an exemplary embodiment, the wind turbine power prediction model further includes a first correction sub-model, which is used to output a first correction parameter for the short-term power prediction result. The wind farm power prediction method based on UAV further includes: extracting actual icing data from the measured icing thickness sequence; inputting the actual icing data into the first correction sub-model to obtain the first correction parameter; and correcting the short-term power prediction result of the wind turbine based on the first correction parameter.

[0015] As an exemplary embodiment, the wind turbine power prediction model further includes a second correction sub-model, which is used to output a second correction parameter for the short-term power prediction result. The method for constructing a wind farm power prediction model based on UAVs includes: extracting each measured static shape feature of icing in the measured icing shape sequence; inputting the static shape features of icing into the second correction sub-model to obtain the second correction parameter; and correcting the short-term power prediction result of the wind turbine based on the second correction parameter.

[0016] This invention provides a method for constructing a wind farm power prediction model based on unmanned aerial vehicles (UAVs). The method includes acquiring historical datasets corresponding to multiple historical icing events at a target wind farm. The historical datasets include multiple historical image sequences and historical power sequences collected by UAVs at preset frequencies. The historical image sequences include multiple historical image data representing the icing state of all wind turbines installed at the target wind farm. The historical image sequences are input into a dual-branch feature extraction model to obtain historical icing thickness sequences and historical icing shape sequences. The dual-branch feature extraction model includes a first edge detection model for extracting icing thickness features and a second edge detection model for extracting icing shape features. The historical icing thickness sequences and historical icing shape sequences are reconstructed to obtain icing state fusion features. The historical icing state fusion feature sequence is then compared with historical data. Temporal dependencies are extracted from historical power sequences to obtain temporal variation features of icing state fusion characteristics related to historical power sequences. A training dataset is constructed based on the temporal variation features and meteorological data. The training dataset is input into a pre-built preset power prediction model for model training to obtain a wind turbine power prediction model. The historical icing state fusion feature sequence can reflect the temporal variation of the wind turbine's icing state, while historical power variation information can reflect the temporal variation of the wind turbine's power. By using the above model training method, the parameters of the preset power prediction model can be continuously adjusted during the training process, so that the obtained wind turbine power prediction model can adapt to the co-occurrence of icing state changes and power changes under icing conditions. In this way, it can predict the co-occurrence of icing changes and output power changes during the icing and melting processes when the wind turbine is in an icing state. Attached Figure Description

[0017] To more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the drawings used in the description of the specific embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0018] Figure 1 This is a flowchart illustrating the method for constructing a wind farm power prediction model based on an unmanned aerial vehicle (UAV) according to an embodiment of the present invention.

[0019] Figure 2 This is a flowchart illustrating a wind farm power prediction method based on unmanned aerial vehicles (UAVs) according to an embodiment of the present invention. Detailed Implementation

[0020] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0021] Wind farms are often located in remote mountainous or forested areas with harsh weather conditions and weak communication signals. As a result, they are prone to icing in extremely cold weather, which is difficult to detect and report in a timely manner. This leads to a significant drop in the accuracy of wind power prediction for a period of time, causing great losses.

[0022] For wind turbine power prediction under icing conditions, related technologies use meteorological conditions such as temperature, humidity, and wind speed corresponding to the wind turbine under icing conditions and their corresponding power to construct a related dataset and train a model. Alternatively, they use the icing thickness corresponding to the wind turbine under icing conditions and its corresponding power to construct a related dataset and train a model. Both of these methods use a one-to-one mapping between meteorological conditions or icing thickness and power. When making power predictions, the output power value at the current moment is usually obtained based on the meteorological conditions or icing thickness at different times.

[0023] However, wind turbine blade icing is a dynamic process, with the shape of the icing blades constantly changing. For example, at the beginning of icing, it takes a certain amount of time for the ice layer to reach a certain thickness. Furthermore, it also takes a certain amount of time for the ice layer to completely melt after reaching a certain thickness, and the melting rate varies at different times. For instance, whether using natural melting or manual de-icing methods, the melting rate is uneven in the morning, noon, and afternoon within the same day. Moreover, within a continuous period, the melting rate is also uneven between two non-contiguous preset time periods. When using a one-to-one mapping method between icing thickness and power for prediction, the power prediction results are inaccurate.

[0024] In view of this, according to embodiments of the present invention, an embodiment of a method for constructing a wind farm power prediction model based on unmanned aerial vehicles is provided. It should be noted that the steps shown in the flowchart in the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions. Furthermore, although a logical order is shown in the flowchart, in some cases, the steps shown or described may be executed in a different order than that shown here.

[0025] This embodiment provides a method for constructing a wind farm power prediction model based on unmanned aerial vehicles (UAVs). Figure 1 This is a flowchart of a method for constructing a wind farm power prediction model based on an unmanned aerial vehicle (UAV) according to an embodiment of the present invention, as shown below. Figure 1 As shown, the process includes the following steps:

[0026] Step S101: Obtain historical datasets corresponding to multiple historical icing events at the target wind farm; wherein, the historical datasets include multiple historical image sequences and historical power sequences collected by drones at preset frequencies; the historical image sequences include multiple historical image data used to represent the icing status of all wind turbines installed in the target wind farm.

[0027] In this invention, the drone conducts inspections of the target wind farm according to a preset frequency and a preset inspection path. The preset frequency is determined based on the area of ​​the target wind farm and the preset inspection path. The area and the preset inspection path can be pre-defined by performing 3D modeling and image recognition of the target wind farm using GIS and remote sensing technologies. Specifically, the drone is controlled to perform an inspection every 5 minutes according to the preset inspection path, and each inspection lasts for 2-3 minutes.

[0028] Monitoring images during the winter icing period are often accompanied by fog interference, and rainy and foggy weather can interfere with image recognition. Therefore, a fog degradation model is established on historical image data collected by UAVs to recover fog-free images.

[0029] The images collected by drones are mostly color images, which contain a huge amount of information. Directly processing color images would slow down the processor. In addition, the most critical factor for object recognition in images is the gradient, and calculating the gradient naturally requires grayscale images. Therefore, historical image data is preprocessed to convert color images into grayscale images.

[0030] Dehazing can improve image quality, but noise issues still exist. Noise is a major interfering factor in image processing, which may occur during image transmission or quantization. Therefore, noise reduction filtering is required to obtain a grayscale image with less noise.

[0031] Step S102: Input the historical image sequence into the dual-branch feature extraction model to obtain the historical ice thickness sequence and the historical ice shape sequence. The dual-branch feature extraction model includes a first edge detection model for extracting ice thickness features and a second edge detection model for extracting ice shape features.

[0032] In this embodiment, the first and second edge detection models can be implemented using image recognition to extract ice thickness and ice shape. As a possible implementation, the Canny operator can be used to obtain ice thickness and ice shape features. The Canny operator is an edge detection algorithm that first smooths the surface and then differentiates, which improves the algorithm's sensitivity to scene edges and suppresses noise. Furthermore, it not only detects fewer false responses but also locates edges that are closest to the true edges.

[0033] While the operator can accurately locate the actual ice-covered edge and shape after edge detection, it is prone to misclassifying noise points as edge points. Therefore, short edges need to be filtered out; for example, an eight-neighborhood tracking algorithm can be used.

[0034] As another possible implementation, deep learning, machine learning, and other methods can be used to extract image features representing the shape information of ice frost from each set of historical image sequences. For example, CNNs can be used for image recognition. In deep image recognition, convolutional neural networks even outperform humans in tasks such as classifying objects into fine-grained categories. In addition, the most popular deep learning models such as YOLO, SSD, and RCNN also use convolutional layers to parse images or photos.

[0035] For example, support vector machines, feature models, etc., can also be used to extract image features representing the ice shape and ice thickness features of each set of historical image sequences.

[0036] Step S103: Reconstruct the features of the historical icing thickness sequence and the historical icing shape sequence to obtain the historical icing state fusion features.

[0037] In this embodiment, to fully represent the icing state, the historical icing thickness sequence and the historical icing shape sequence features are reconstructed to obtain the historical icing state fusion features. As one possible implementation, the historical icing thickness sequence and the historical icing shape sequence can be directly added together for reconstruction. As another possible implementation, the historical icing thickness sequence and the historical icing shape sequence can be assigned corresponding weights according to their own attributes, and then weighted and fused based on their weights for feature reconstruction. Among them, the fusion weight of the historical icing thickness sequence is proportional to the detection accuracy of the first edge detection model, the fusion weight of the historical icing shape sequence is proportional to the detection accuracy of the second edge detection model, and the sum of the fusion weights of the historical icing thickness sequence and the historical icing shape sequence is 1.

[0038] Step S104: Extract the temporal dependency between the historical icing state fusion feature sequence and the historical power sequence to obtain the historical temporal change features of the historical icing state fusion features related to the historical power sequence.

[0039] To determine the dynamic changes in icing shape and thickness, this embodiment extracts the temporal dependencies of historical icing thickness sequences, historical icing shape sequences, and historical power sequences. This yields historical icing thickness change information, historical icing shape change information, and historical power change information as historical temporal change features of the historical icing state fusion features related to the historical power sequence. As a possible implementation, the historical icing thickness sequence and historical power sequence can be input into a Gated Recurrent Unit (GRU) or a Long Short-Term Memory (LSTM) network to extract the temporal dependencies of the historical icing thickness sequence and historical power sequence through GRU or LSTM, thereby obtaining historical icing thickness change information, historical icing shape change information, and historical power change information, respectively.

[0040] Step S105: Construct a training dataset based on historical time-series change characteristics and meteorological data, and input the training dataset into a pre-constructed preset power prediction model for model training to obtain a wind turbine power prediction model.

[0041] In this embodiment, the input to the pre-built preset power prediction model may include multi-dimensional data; for example, the input to the preset power prediction model may include meteorological data, icing thickness data, etc., wherein the meteorological data may include temperature, humidity, wind speed, pressure, etc.; based on this, a training dataset is constructed based on historical time-series variation characteristics and meteorological data; specifically, a training dataset can be constructed based on historical icing thickness variation information, historical icing shape variation information, historical power variation information, meteorological data, and historical time-series variation information corresponding to each icing event. After constructing the dataset, the training data is input into the pre-built preset power prediction model for model training. During the training process, the time-series correspondence between historical icing thickness variation information, historical icing shape variation information, historical time-series variation information, and historical power variation information is learned to obtain the wind turbine power prediction model.

[0042] In the above implementation method of this embodiment, the historical icing state fusion feature sequence can reflect the temporal changes in the icing state of the wind turbine, while the historical power change information can reflect the temporal changes in the power of the wind turbine. By using the above model training method, the parameters of the preset power prediction model can be continuously adjusted during the training process, so that the obtained wind turbine power prediction model can adapt to the accompanying relationship between icing state changes and power changes under icing conditions, and thus can predict the accompanying relationship between icing changes and output power changes during the icing process and the melting process when the wind turbine is in an icing state.

[0043] As an exemplary embodiment, the wind turbine power prediction model includes at least two wind turbine power prediction sub-models, which are used for power prediction at different historical icing change rates. The method for constructing a wind farm power prediction model based on UAVs further includes: truncating the historical icing state fusion feature sequence according to a preset time period to obtain a historical icing state fusion feature sub-sequence; clustering the historical icing state fusion feature sub-sequence according to the historical icing change rate to obtain at least two historical clusters; and dividing the training dataset according to the historical clusters to obtain at least two training data subsets corresponding to different historical icing change rates.

[0044] Different wind turbine power prediction sub-models were trained using subsets of training data to obtain wind turbine power prediction sub-models corresponding to different historical icing change rates, so as to predict the power of input data corresponding to different historical icing change rates.

[0045] Temporal dependencies were extracted from the fusion feature sequences of historical icing states for each historical cluster and the corresponding historical icing state, respectively, to obtain the training dataset for each historical cluster. The training dataset includes information on historical icing thickness variation, historical icing shape variation, and historical power variation for each historical cluster. The training dataset was then input into a pre-built preset power prediction model for model training to obtain the short-term power prediction sub-model for each historical cluster.

[0046] Different historical icing change rates under icing conditions will result in different icing thickness and icing shape of wind turbine blades per unit time, which in turn will lead to different output power of the wind turbine. Therefore, in order to distinguish the impact of different historical icing change rates on power, historical clusters are divided according to different historical icing change rates, so that the wind turbine power prediction model can adapt to different historical icing change rates for power prediction.

[0047] In this embodiment, the historical icing state fusion feature sequence is first truncated according to a preset time period to obtain a historical icing state fusion feature subsequence. After obtaining the historical icing state fusion feature subsequence, K-clustering can be used to cluster the historical icing state fusion feature subsequence according to the historical icing change rate to obtain at least two historical clusters. Each historical cluster can reflect the icing thickness change, icing shape change, and power change sequence corresponding to different historical icing change rates. Further, the training dataset is divided according to the historical clusters to obtain at least two training data subsets corresponding to different historical icing change rates.

[0048] In this embodiment, GRU and LSTM can be used to extract the temporal dependencies of each historical cluster and the corresponding historical icing state fusion feature subsequence, respectively, to obtain the training dataset corresponding to each historical cluster. The training dataset includes historical icing thickness change information, historical icing shape change information, and historical power change information corresponding to each historical cluster. Different wind turbine power prediction sub-models are trained using the subset of training data to obtain wind turbine power prediction sub-models corresponding to different historical icing change rates, so as to predict the power of the input data corresponding to different historical icing change rates.

[0049] As an exemplary embodiment, the historical icing change rate includes the icing thickness change rate and / or icing shape change rate.

[0050] In this invention, in order to simultaneously consider the impact of icing thickness on short-term power prediction results from the perspective of static features, and thus enable the trained model to correct the short-term power prediction results based on these static features, as an exemplary embodiment, the wind turbine power prediction model further includes a first correction sub-model. The first correction sub-model is used to output the first correction parameter of the short-term power prediction result. The method for constructing a wind farm power prediction model based on UAVs includes: extracting each historical icing thickness value and the historical power data corresponding to the time sequence of the historical icing thickness values ​​from the historical icing thickness sequence as a first training set; inputting the first training set into the first correction sub-model for model training, and learning the correspondence between historical icing thickness values, historical power data and the first correction parameter during the training process.

[0051] In this embodiment, the first correction sub-model is used to output the first correction parameter of the power prediction result. The first correction parameter can be a correction coefficient, a correction value, or a combination of a correction coefficient and a correction value.

[0052] For example, after obtaining the historical icing thickness sequence, each historical icing thickness value and the historical power data corresponding to the time series of the historical icing thickness values ​​are extracted from the historical icing thickness sequence as a first training set. The first training set is then input into a first correction sub-model for model training. The first correction sub-model can be a neural network model, which maps the mathematical relationship between historical icing thickness values, time-series historical power data, and first correction parameters. Specifically, the training process involves inputting the first training set into the first correction sub-model to obtain the first correction parameters output by the first correction sub-model. These first correction parameters are then mathematically calculated with the short-term wind power prediction result output by the trained first wind power short-term power prediction model to obtain the first corrected prediction result. In each iteration, the parameters of the first correction sub-model are continuously adjusted to minimize the error between the first corrected prediction result and the historical power data corresponding to the time series of the first corrected prediction result, until the error meets a preset condition.

[0053] When a wind turbine is iced, ice covers the blade surface and its shape gradually changes. This shape change is like constantly replacing the blades, and the shape of the replaced blades is difficult to determine accurately, resulting in continuously deteriorating aerodynamic performance and altered output power. Therefore, changes in ice shape have a significant impact on the wind turbine's output power. Based on this, as an exemplary embodiment, the wind turbine power prediction model also includes a second correction sub-model. The second correction sub-model is used to output second correction parameters for short-term power prediction results. The method for constructing a wind farm power prediction model based on UAVs includes: extracting historical static shape features of each historical icing shape sequence and historical power data corresponding to the time sequence of historical static shape features of historical icing as a second training set; inputting the second training set into the second correction sub-model for model training, and learning the correspondence between historical static shape features of icing, historical power data, and second correction parameters during the training process.

[0054] In this embodiment, the second correction sub-model is used to output the second correction parameter of the short-term power prediction result. The second correction parameter can be a correction coefficient, a correction value, or a combination of a correction coefficient and a correction value.

[0055] For example, the historical static shape features of each historical icing shape in the historical icing shape sequence and the historical power data corresponding to the time sequence of the historical static shape features of the historical icing are extracted as the second training set.

[0056] In this embodiment, the historical icing static shape features may include the cross-sectional static shape of the icing and the surface interface static shape.

[0057] For example, the temporal dependency of the historical icing static shape features is extracted to obtain the historical icing static shape features; for details, please refer to the content described in the above embodiments, which will not be repeated here.

[0058] Further, historical power data corresponding to the time series of historical icing static shape features is obtained as a second training dataset. The second training dataset is input into the second correction sub-model. During training, the correspondence between historical icing static shape features, historical power data, and the second correction parameter is learned to obtain the second correction sub-model. The second correction sub-model can be a neural network model, which is used to map the mathematical relationship between historical icing static shape features, time series-corresponding historical power data, and the second correction parameter. Specifically, the training process can be as follows: the second training set is input into the second correction sub-model to obtain the second correction parameter output by the second correction sub-model. The second correction parameter is then mathematically calculated with the wind power short-term power prediction result output by the trained first wind power short-term power prediction model or the second wind power short-term power prediction model to obtain the second correction prediction result. In each iteration, the parameters of the second correction sub-model are continuously adjusted to make the error between the second correction prediction result and the historical power data corresponding to the time series of the second correction prediction result smaller, until the error meets the preset condition.

[0059] Secondly, this invention provides a wind farm power prediction method based on unmanned aerial vehicles (UAVs). Figure 2 This is a flowchart of a wind farm power prediction method based on an unmanned aerial vehicle (UAV) according to an embodiment of the present invention, as shown below. Figure 2 As shown, the process includes the following steps:

[0060] Step S201: Obtain the measured image sequence and predicted meteorological data of the target wind farm. The measured image sequence is obtained by the UAV according to the preset acquisition frequency.

[0061] In this embodiment, the measured image sequence is acquired by a drone; as a possible implementation, the drone can be controlled to perform an inspection every 5 minutes according to a preset inspection path, with each inspection lasting 2-3 minutes.

[0062] In this embodiment, the predicted meteorological data can be determined through numerical weather prediction.

[0063] As an exemplary embodiment, the time length of the measured image sequence and the time length of the predicted meteorological data can be adaptively selected according to the required short-term power prediction results of the wind turbine.

[0064] Step S202: Input the measured image sequence into the dual-branch feature extraction model to obtain the measured ice thickness sequence and the measured ice shape sequence. The dual-branch feature extraction model includes a first edge detection model for extracting ice thickness features and a second edge detection model for extracting ice shape features.

[0065] In this embodiment, the dual-branch feature extraction model can extract the ice thickness and ice shape of the measured image sequence through image recognition, thereby obtaining the measured ice thickness sequence and the measured ice shape sequence; the specific implementation method can be found in the content described in the above embodiment, and will not be repeated here.

[0066] Step S203: Reconstruct the features of the measured ice thickness sequence and the measured ice shape sequence to obtain the measured ice state fusion features.

[0067] In this embodiment, in order to fully represent the icing state, the measured icing thickness sequence and the measured icing shape sequence features are reconstructed to obtain the measured icing state fusion features; the specific implementation method can be found in the content described in the above embodiment, and will not be repeated here.

[0068] Step S204: Extract the temporal dependency of the measured icing state fusion feature sequence to obtain the actual temporal change features of the measured icing state fusion features;

[0069] In order to determine the temporal variation of the ice thickness of the wind turbine corresponding to the current measured image sequence, this embodiment extracts the temporal dependency of the measured ice thickness sequence; the specific implementation method can be found in the content described in the above embodiment, and will not be repeated here.

[0070] Step S205: Based on the actual time-series variation characteristics and measured meteorological data, input the data into the pre-trained wind turbine power prediction model to perform power prediction and obtain the wind farm power prediction results.

[0071] In this embodiment, the wind turbine power prediction model is constructed based on historical icing thickness variation information, historical icing shape variation information, historical power variation information, meteorological data, and historical time-series variation information corresponding to each icing event. The training dataset is input into a pre-constructed preset power prediction model for model training. During the training process, the model learns the time-series correspondence between historical icing thickness variation information, historical icing shape variation information, historical time-series variation information, and historical power variation information. This training method can consider the co-occurrence relationship between icing thickness variation and power variation under icing conditions, and thus can predict the co-occurrence relationship between the output power change during the icing process and the ice thickness change and output power change during the melting process when the wind turbine is in an icing state. This can solve the problem that related technologies do not consider the impact of icing conditions on power prediction over time, resulting in inaccurate power prediction under icing conditions.

[0072] Different historical icing change rates under icing conditions lead to varying icing thicknesses on wind turbine blades per unit time, resulting in different output power. Therefore, in this invention, to distinguish the impact of different historical icing change rates on power, the icing state fusion features and the blade surface are clustered for different time periods. As an exemplary embodiment, the wind turbine power prediction model includes at least two wind turbine power prediction sub-models, each used for power prediction at different icing thickness change rates. The UAV-based wind farm power prediction method further includes: dividing the measured icing state fusion feature sequence according to a preset duration to obtain a measured icing state fusion feature sub-sequence; clustering the measured icing state fusion feature sub-sequence and historical clusters according to the actual icing change rate to obtain clustering results; based on the clustering results, selecting the wind turbine power prediction sub-model corresponding to a historical cluster with a clustering degree greater than a preset clustering degree, and inputting the measured icing state fusion feature sub-sequence into the wind turbine power prediction sub-model to obtain the corresponding wind turbine power prediction sub-result.

[0073] In this embodiment, the wind turbine power prediction sub-model is obtained by training the model after fusing the feature sequences of historical icing states according to the actual icing change rate. Therefore, selecting the short-term power prediction sub-model of the wind turbine corresponding to the cluster with a cluster degree greater than the preset degree for power prediction can take into account the different icing thickness of the wind turbine blades per unit time caused by different historical icing change rates under icing conditions, which in turn leads to different output power of the wind turbine, making it more accurate.

[0074] For example, the measured icing state fusion features are truncated according to a preset duration to obtain a measured icing state fusion feature subsequence. After obtaining the measured icing state fusion feature subsequence, clustering is performed based on the measured icing state fusion feature subsequence and clusters. The short-term power prediction sub-model of the wind turbine corresponding to the cluster with a cluster degree greater than a preset degree is selected. The above process can be implemented using the K-clustering method. After obtaining the clustering results, the short-term power prediction sub-model of the wind turbine corresponding to the cluster with a cluster degree greater than a preset degree is selected. The measured icing state fusion feature subsequence is input into the short-term power prediction sub-model of the wind turbine to obtain the short-term power prediction result of the wind turbine.

[0075] As an exemplary embodiment, the actual icing change rate includes the icing thickness change rate and / or icing shape change rate.

[0076] To simultaneously consider the impact of icing thickness on short-term power prediction results from the perspective of static features, as an exemplary embodiment, the wind turbine power prediction model further includes a first correction sub-model. The first correction sub-model is used to output a first correction parameter for the short-term power prediction results. The wind farm power prediction method based on UAVs further includes: extracting actual icing data from the measured icing thickness sequence; inputting the actual icing data into the first correction sub-model to obtain the first correction parameter; and correcting the short-term power prediction results of the wind turbine based on the first correction parameter.

[0077] In this embodiment, the first correction sub-model is trained using historical icing thickness values ​​and historical power data corresponding to the time series of historical icing thickness values ​​in the historical icing thickness sequence as the first training set. The first correction sub-model is used to learn the correspondence between historical icing thickness values, historical power data and the first correction parameter. Based on this, the actual icing data is input into the first correction sub-model to obtain the first correction parameter that considers the impact of icing thickness on short-term power prediction results from the perspective of static features, and the short-term power prediction results are corrected using the first correction parameter.

[0078] As an exemplary embodiment, the first correction sub-model is used to output a first correction parameter for the short-term power prediction result. The first correction parameter can be a correction coefficient, a correction value, or a combination of a correction coefficient and a correction value.

[0079] In the above embodiments of this application, the actual icing data in the measured icing thickness sequence can reflect the static situation of icing thickness. Therefore, by adopting the above power prediction method, the influence of the static situation of icing thickness on power prediction can be considered in the power prediction process.

[0080] As an exemplary embodiment, the wind farm power prediction method based on UAVs further includes: extracting each measured static shape feature of icing in the measured icing shape sequence; inputting the measured static shape features of icing into a second correction sub-model to obtain a second correction parameter; and correcting the short-term power prediction result based on the second correction parameter.

[0081] In this embodiment, image recognition can be used to extract the ice-covered shape from the measured image sequence to obtain the measured static shape features of the ice-covered area; the specific implementation method can be referred to the content described in the above embodiments, and will not be repeated in this embodiment.

[0082] In this embodiment, the second correction sub-model is trained using the historical icing shape features in the historical icing shape sequence and the historical power data corresponding to the time sequence of the historical icing shape features as the second training set. The second correction sub-model is used to learn the correspondence between the historical icing static shape features, the historical power data and the second correction parameter. Based on this, by inputting the historical icing static shape features into the second correction sub-model, the second correction parameter that considers the influence of icing shape on the short-term power prediction result from the perspective of static features can be obtained, and the short-term power prediction result is corrected using the second correction parameter.

[0083] In the above-described embodiments of this application, the actual icing data in the measured icing thickness sequence can reflect the static condition of the icing shape. Therefore, by adopting the above-described power prediction method, the influence of the static condition of the icing shape on power prediction can be considered during the power prediction process.

[0084] As an exemplary embodiment, the second correction sub-model is used to output a second correction parameter for the short-term power prediction result. The second correction parameter can be a correction coefficient, a correction value, or a combination of a correction coefficient and a correction value.

[0085] As an exemplary embodiment, the measured static shape characteristics of the ice accretion may include the cross-sectional static shape of the ice accretion and the surface interface static shape.

[0086] The sequence numbers of the embodiments in this application are for descriptive purposes only and do not represent the superiority or inferiority of the embodiments.

[0087] The above are merely preferred embodiments of this application. It should be noted that those skilled in the art can make various improvements and modifications without departing from the principles of this application, and these improvements and modifications should also be considered within the scope of protection of this application.

[0088] In the above embodiments of this application, the descriptions of each embodiment have different focuses. For parts not described in detail in a certain embodiment, please refer to the relevant descriptions of other embodiments.

[0089] The above are merely preferred embodiments of this application. It should be noted that those skilled in the art can make various improvements and modifications without departing from the principles of this application, and these improvements and modifications should also be considered within the scope of protection of this application.

Claims

1. A method for constructing a wind farm power prediction model based on unmanned aerial vehicles (UAVs), characterized in that, The method for constructing a wind farm power prediction model based on unmanned aerial vehicles (UAVs) includes: Obtain historical datasets corresponding to multiple icing events at the target wind farm; wherein, the historical datasets include multiple historical image sequences and historical power sequences collected by drones at preset frequencies; The historical image sequence includes multiple historical image data representing the icing status of all wind turbines installed within the target wind farm. The historical image sequence is input into a dual-branch feature extraction model to obtain a historical ice thickness sequence and a historical ice shape sequence. The dual-branch feature extraction model includes a first edge detection model for extracting ice thickness features and a second edge detection model for extracting ice shape features. The historical icing thickness sequence and historical icing shape sequence features are reconstructed to obtain icing state fusion features; Temporal dependency is extracted between the icing state fusion feature sequence and the historical power sequence to obtain the temporal variation features of the icing state fusion features related to the historical power sequence; A training dataset is constructed based on the time-series variation characteristics and meteorological data. The training dataset is then input into a pre-constructed preset power prediction model for model training to obtain a wind turbine power prediction model. The wind turbine power prediction model includes at least two wind turbine power prediction sub-models, each used for power prediction at different icing change rates. The method for constructing the UAV-based wind farm power prediction model also includes: The icing state fusion feature sequence is divided according to a preset duration to obtain an icing state fusion feature subsequence; The fused feature subsequences of the icing state are clustered according to the icing change rate to obtain at least two historical clusters; The training dataset is divided according to the historical clusters to obtain at least two training data subsets corresponding to different icing change rates; Different wind turbine power prediction sub-models are trained using subsets of training data to obtain wind turbine power prediction sub-models corresponding to different icing change rates, so as to predict power for input data corresponding to different icing change rates. The icing change rate includes the icing thickness change rate and / or icing shape change rate.

2. The method for constructing a wind farm power prediction model based on unmanned aerial vehicles as described in claim 1, characterized in that, The wind turbine power prediction model further includes a first correction sub-model, which is used to output the first correction parameter of the short-term power prediction result. The method for constructing the wind farm power prediction model based on UAV includes: Extract each historical icing thickness value from the historical icing thickness sequence and the historical power data corresponding to the time series of the historical icing thickness values ​​as the first training set; The first training set is input into the first correction sub-model for model training. During the training process, the correspondence between the historical icing thickness value, the historical power data and the first correction parameter is learned.

3. The method for constructing a wind farm power prediction model based on unmanned aerial vehicles as described in claim 1 or 2, characterized in that, The wind turbine power prediction model further includes a second correction sub-model, which is used to output a second correction parameter for the short-term power prediction result. The method for constructing the wind farm power prediction model based on UAV includes: Extract each historical static shape feature of icing in the historical icing shape sequence and the historical power data corresponding to the time sequence of the historical static shape features of icing as the second training set; The second training set is input into the second correction sub-model for model training. During the training process, the correspondence between the historical icing static shape features, the historical power data and the second correction parameters is learned.

4. A wind farm power prediction method based on unmanned aerial vehicles (UAVs), characterized in that, The UAV-based wind farm power prediction method includes: Acquire a sequence of measured images and predicted meteorological data of the target wind farm, wherein the sequence of measured images is acquired by a UAV at a preset acquisition frequency; The measured image sequence is input into a dual-branch feature extraction model to obtain a measured ice thickness sequence and a measured ice shape sequence. The dual-branch feature extraction model includes a first edge detection model for extracting ice thickness features and a second edge detection model for extracting ice shape features. The measured ice thickness sequence and the measured ice shape sequence features are reconstructed to obtain the measured ice state fusion features. Temporal dependencies were extracted from the measured icing state fusion feature sequence to obtain the actual temporal variation features of the measured icing state fusion features; Based on the actual time-series variation characteristics and measured meteorological data, the power prediction is performed by inputting the pre-trained wind turbine power prediction model to obtain the wind farm power prediction results. The wind turbine power prediction model includes at least two wind turbine power prediction sub-models, each used for power prediction at different rates of icing thickness variation. The UAV-based wind farm power prediction method further includes: The measured icing state fusion feature sequence is divided according to a preset duration to obtain the measured icing state fusion feature subsequence; The measured icing state fusion feature subsequence and historical clusters are clustered according to the actual icing change rate to obtain the clustering results; Based on the clustering results, select the wind turbine power prediction sub-model corresponding to the historical clusters with a clustering degree greater than the preset clustering degree, and input the measured icing state fusion feature sub-sequence into the wind turbine power prediction sub-model to obtain the corresponding wind turbine power prediction sub-result; the actual icing change rate includes the icing thickness change rate and / or icing shape change rate.

5. The wind farm power prediction method based on UAV as described in claim 4, characterized in that, The wind turbine power prediction model also includes a first correction sub-model, which is used to output the first correction parameter of the short-term power prediction result. The wind farm power prediction method based on UAV also includes: Extract actual icing data from the measured icing thickness sequence; The actual icing data is input into the first correction sub-model to obtain the first correction parameter; The short-term power prediction results of the wind turbine are corrected based on the first correction parameter.

6. The wind farm power prediction method based on UAV as described in claim 4 or 5, characterized in that, The wind turbine power prediction model further includes a second correction sub-model, which is used to output a second correction parameter for the short-term power prediction result. The method for constructing the wind farm power prediction model based on UAV includes: Extract the static shape features of each measured ice accretion from the measured ice accretion shape sequence; input the static shape features of the ice accretion into the second correction sub-model to obtain the second correction parameters; The short-term power prediction results of the wind turbine are corrected based on the second correction parameter.

Images