A wind power prediction method and system integrating three-dimensional convolution and lightweight convolution threshold unit
By integrating wind power prediction methods with three-dimensional convolution and lightweight convolution threshold units, the problem of low prediction accuracy in the existing technology under extreme meteorological conditions is solved, and a higher precision wind power prediction is achieved, suitable for wind power on multiple time scales Power prediction.
Patent Information
- Application Number
- CN202111060872.3
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2021-09-10
- Publication Date
- 2024-11-05
- Estimated Expiration
- 2041-09-10
Smart Images

Figure CN113962432B_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of wind power prediction, and in particular to a wind power prediction method and system integrating three-dimensional convolution and lightweight convolution threshold units. Background Art
[0002] With the vigorous promotion of renewable energy power generation, and in response to my country's policy of encouraging renewable energy to access the power grid first, the research on wind power generation has received widespread attention. Due to the randomness, intermittency and uncertainty of wind energy, it has an adverse effect on the safe and stable operation of the power grid. If the wind power can be predicted more effectively and accurately, it can not only provide valuable data support for the power sector, but also fundamentally alleviate or even solve the safety problems caused by wind power grid connection to the power grid. High-precision prediction of wind power under extreme meteorological conditions is of great significance to improving the dynamic performance of wind turbines and maintaining the safe and stable operation of the power grid.
[0003] At present, the research on wind power prediction has become more and more extensive. Wind power prediction methods can be mainly divided into three categories: physical methods, statistical methods, and learning methods. The physical method mainly constructs a physical prediction model based on the principles of atmospheric motion and fluid mechanics. The model construction and solution process of this method is relatively complex, and it is easy to produce error accumulation effects, thereby affecting the prediction accuracy. The statistical method uses curve fitting, parameter estimation and other methods to make predictions based on historical wind speed or wind power data. However, the statistical characteristics of the historical data itself will cause the prediction error to gradually increase as the prediction time increases, and it is not sensitive enough to sudden situations. Therefore, it is not suitable for wind power prediction under extreme weather conditions. Compared with statistical methods, although traditional machine learning methods require a large amount of historical data as input, they do not need to establish a specific mathematical relationship between input and output objects, but they are prone to fall into the local optimum of the model or overfitting problems, resulting in reduced prediction accuracy. Summary of the invention
[0004] The purpose of the present invention is to provide a wind power prediction method and system that integrates three-dimensional convolution and lightweight convolution threshold units, and predicts wind power by fully learning the multi-dimensional attributes of wind turbines and meteorological forecast information, avoiding large deviations under extreme meteorological conditions and improving prediction accuracy.
[0005] To achieve the above object, the technical solution adopted by the present invention is as follows:
[0006] The present invention provides a wind power prediction method integrating three-dimensional convolution and lightweight convolution threshold unit, comprising:
[0007] Establish a wind power prediction model that integrates three-dimensional convolution and lightweight convolution threshold units;
[0008] The wind power prediction model established by collecting historical spatiotemporal data of wind farm units for training;
[0009] The real-time spatiotemporal data of wind farm units are input into the trained wind power prediction model to obtain the power prediction results of wind farm units.
[0010] Furthermore, the spatiotemporal data of wind farm units include: endogenous attribute data and external state data;
[0011] The endogenous attribute data includes cut-in wind speed, cut-out wind speed and control strategy matrix;
[0012] The external state data includes wind farm unit power, wind farm wind speed and weather forecast wind speed.
[0013] Furthermore, it also includes:
[0014] The nearest neighbor linear interpolation algorithm is used to interpolate missing values and outliers in the spatiotemporal data of wind farm units.
[0015] Furthermore, it also includes:
[0016] The processed spatiotemporal data of wind farm units are standardized to limit the data values to the interval [0, 1].
[0017] Furthermore, it also includes:
[0018] The spatiotemporal data of wind farm units are divided into short-term series X S and the medium- and long-term series X L , expressed as:
[0019]
[0020]
[0021] Among them, the short-term series X S It means taking the current time point t as the benchmark, taking 1 sampling unit as the step length, and taking d data in sequence; the medium and long-term series X L It means taking the current time point t as the reference, taking q sampling units as the step length, and taking c data forward in sequence; X t represents the spatiotemporal data at time point t, It represents the j-dimensional spatiotemporal data of the i-th unit object at time t, N is the number of wind farm units, M is the spatiotemporal data dimension of wind farm units, T is the total number of data in short-term series and medium- and long-term series, and satisfies T=d+c.
[0022] Furthermore, the wind power prediction model integrating three-dimensional convolution and lightweight convolution threshold unit is established, including:
[0023] Two 3D convolutional modules are stacked to construct a 3D convolutional network to train the short-term sequence X S Capture spatiotemporal features for learning input;
[0024] A double-layer lightweight convolution threshold unit is used to store medium- and long-term sequences X L To learn and capture spatiotemporal features for input; the lightweight convolution threshold unit is expressed as:
[0025]
[0026] Among them, * represents the convolution operation, represents the Hadamard product, R t , Z t , H t and H′ t Represent the reset gate, update gate, current memory state and candidate memory state respectively, C t The input three-dimensional tensor is the medium- and long-term sequence X L , W HZ , W CR , W HR and W HH is the corresponding two-dimensional convolution kernel, B Z , B R and B H is the corresponding bias, σ is the Sigmoid activation function;
[0027] A fully connected network is used to fuse the output temporal and spatial features, and finally a complete wind farm unit wind power prediction result is obtained.
[0028] Furthermore, the output spatiotemporal features are fused using a fully connected network, including:
[0029] Flatten the captured spatiotemporal features into one-dimensional vectors and concatenate them.
[0030] Use the fully connected layer to map the learned spatiotemporal features to the sample label space, and reconstruct the two spatiotemporal features into a complete feature map through the weight matrix.
[0031] Furthermore, the wind power prediction model is trained using the Adam optimization algorithm.
[0032] Furthermore, during the training process, the mean square error is used as the loss function.
[0033] The present invention also provides a wind power prediction system integrating three-dimensional convolution and lightweight convolution threshold unit, comprising:
[0034] Long-term feature learning module, used to build a three-dimensional convolutional network to learn the long-term spatiotemporal data of wind farm units;
[0035] The short-term feature learning module is used to construct a double-layer lightweight convolution threshold unit to learn the short-term spatiotemporal data of wind farm units;
[0036] The feature fusion output module is used to fuse the learning results of the long-term feature learning module and the short-term feature learning module using a fully connected network to obtain a complete wind farm unit wind power prediction result.
[0037] Furthermore, the three-dimensional convolutional network is constructed by stacking two three-dimensional convolutional modules.
[0038] Furthermore, the lightweight convolution threshold unit is specifically:
[0039]
[0040] Among them, * represents the convolution operation, represents the Hadamard product, R t , Z t , H t and H′ t Represent the reset gate, update gate, current memory state and candidate memory state respectively, C t The input three-dimensional tensor is the medium- and long-term sequence X L , W HZ , W CR , W HR and W HH is the corresponding two-dimensional convolution kernel, B Z , B R and B H is the corresponding bias, and σ is the Sigmoid activation function.
[0041] Furthermore, it also includes:
[0042] The training module is used to train the long-term feature learning module, the short-term feature learning module and the feature fusion output module based on the historical spatiotemporal data of the wind farm units.
[0043] Compared with the prior art, the present invention has the following advantages:
[0044] (1) The power generation of wind farm units is affected by the monsoon climate and shows seasonal periodic changes. The present invention has objective advantages in using multi-scale time observation data to predict wind power; (2) When wind energy propagates between wind turbines, each wind turbine affects each other. Multiple prediction tasks with strong correlation can exchange information, and multi-target output prediction is a more effective solution; (3) Considering the complexity and diversity of wind power spatiotemporal data, the multi-channel fusion method is used to learn the spatiotemporal correlation features, which can enhance the comprehensive extraction capability of data features and improve the prediction accuracy; (4) Compared with traditional machine learning methods, the deep learning method can extract and retain data features more comprehensively, and its prediction model has better adaptability; in addition, the deep learning method does not need to establish a specific analytical relationship between the predicted power and meteorological data. Good prediction results can be obtained through a large amount of data training, which is suitable for ultra-short-term, short-term and medium- and long-term power forecasting. BRIEF DESCRIPTION OF THE DRAWINGS
[0045] Figure 1 A schematic diagram of a flow chart of a wind power prediction method integrating three-dimensional convolution and lightweight convolution threshold unit provided by the present invention;
[0046] Figure 2 The present invention provides a wind power prediction system architecture that integrates three-dimensional convolution and lightweight convolution threshold units. DETAILED DESCRIPTION
[0047] The present invention is further described below. The following examples are only used to more clearly illustrate the technical solution of the present invention, and are not intended to limit the protection scope of the present invention.
[0048] The embodiment of the present invention provides a wind power prediction method integrating three-dimensional convolution and lightweight convolution threshold unit, see Figure 1 , as follows:
[0049] Step 1: obtain the spatiotemporal data information of the wind farm, including the endogenous attribute data and external state data of each wind farm unit, and divide it into training samples and test samples;
[0050] Step 2, matrix representation and multi-time scale division of the preprocessed sample data;
[0051] Step 3: Establish a wind power prediction model that integrates three-dimensional convolution and lightweight convolution threshold unit. Use three-dimensional convolution network and lightweight convolution threshold unit as parallel modules to respectively mine the short-term fluctuation and long-term trend characteristics of the spatiotemporal data of wind farm units. Finally, use the fully connected layer to fuse the multi-scale features.
[0052] Step 4: training and optimizing the parameters of the wind power prediction model through training samples;
[0053] Step 5: Use the trained wind power prediction model to perform multi-unit wind power prediction on the test samples.
[0054] See also Figure 2 In an embodiment of the present invention, the acquired wind farm sample data includes endogenous attribute data and external state data of each wind farm unit, wherein the endogenous attribute data includes cut-in wind speed, cut-out wind speed and control strategy matrix; the external state data includes historical power, historical wind speed and meteorological forecast wind speed.
[0055] The nearest neighbor linear interpolation algorithm is used to interpolate missing values and outliers. At the same time, due to differences in data dimensions and value ranges, in order to speed up model training and improve prediction results, standardization is performed to limit data values to the interval [0, 1]. The sample data is then divided into training samples and test samples.
[0056] See also Figure 2 In the embodiment of the present invention, the pre-processed sample data is represented by a matrix and divided into multiple time scales. The specific operations are as follows:
[0057] Assume that the source of spatiotemporal data of the wind farm is N units A 1 ,...,A N , the time series length of each object is T, and the generated data dimension is M. When M = 1, it means that the sequence of each object is a vector of length T. At any time t∈{1,...,T}, the spatial data composed of each unit object can be represented as an N×M-dimensional matrix
[0058]
[0059] in, Represents the j-th dimension data of the i-th unit object at time t.
[0060] The spatiotemporal data generated by N units of objects at a continuous time length T can be expressed as a three-dimensional tensor To express.
[0061] X={X 1 ,X 2 ,...,X T}
[0062] If we want to predict the state X of the N units at the target time t+τ t+τ , considering the differences in steady-state and dynamic characteristics of spatiotemporal data at different time scales, and the fact that correlation usually weakens with the increase of time series length, it is necessary to conduct multi-scale joint learning. Without loss of generality, taking two time scales as an example, the spatiotemporal data X input to the prediction model is divided into two parts at different time scales: the near short-term observation sequence X Sand the medium- and long-term observation series X L , and other time scale situations can be expanded from it.
[0063]
[0064] Short-term series X S It means taking the current time point t as the reference and taking 1 as the step length, taking d data forward in sequence as the short-term data. It should be noted that the step length 1 here refers to 1 sampling unit.
[0065] Medium and long term series X L It means taking the current time point t as the benchmark and the step length q as the step length, c data are taken forward in sequence as the medium- and long-term data. It should be noted that the step length q here refers to q sampling units.
[0066] d and c are short-term series X S and the medium- and long-term series X L The length of the sequence, q is the observation step length of the medium and long term sequence.
[0067] See also Figure 2 In the embodiment of the present invention, in order to fully mine the spatiotemporal data of wind turbines, a wind power prediction model integrating three-dimensional convolution and lightweight convolution threshold units is established, including three modules: a long-term feature learning module, a short-term feature learning module and a feature fusion output module.
[0068] (1) The long-term feature learning module uses a lightweight convolution threshold unit to reduce the complexity of network training and improve the time series modeling capability while capturing the long-term trend characteristics of wind turbine data and the temporal and spatial correlation characteristics. Its working mechanism is as follows:
[0069]
[0070] In the above formula, * represents the convolution operation, represents the Hadamard product. t , Z t , H t and H′ t Represent the reset gate, update gate, current memory state and candidate memory state respectively, C t is the input three-dimensional tensor, W HZ , W CR , W HR and W HH is the corresponding two-dimensional convolution kernel, B Z , B R and B H is the corresponding bias, and σ is the Sigmoid activation function.
[0071] The present invention constructs a learning channel of a double-layer lightweight convolution threshold unit to learn long-term scale data, and its operation can be described as:
[0072] X' L =f L (X L )
[0073] In the formula, f L is the nonlinear function corresponding to the learning channel, X L is a medium- to long-term series, X' L is the spatiotemporal features captured through this channel.
[0074] (2) The short-term feature learning module is based on a three-dimensional convolutional network, which is specifically constructed by stacking two streamlined three-dimensional convolutional modules, Conv-1 (Conv3D-1, ReLU1) and Conv-2 (Conv3D-2, ReLU2). It can share the spatial information of related objects, learn the temporal correlation, and effectively extract the small-scale dynamic characteristics of adjacent short-term data. The learning process of adjacent short-term spatiotemporal data is described as:
[0075] X′ S =f S (X S )
[0076] In the formula, f S is the nonlinear function corresponding to the learning channel, X S is the short-term sequence of input, X′ S is the spatiotemporal features captured through this channel.
[0077] (3) The feature fusion output module uses a fully connected network to fuse the spatiotemporal features output by (1) and (2). First, the two-channel features are flattened into a one-dimensional vector and concatenated. The learned features are mapped to the sample label space using a fully connected layer. The local features are reconstructed into a complete feature map through a weight matrix to reduce the interference caused by the relative position of the features on the prediction. The number of neurons in the last fully connected layer is set to the number of predicted objects, so that multi-object prediction can be achieved. The output module can be described by the following formula:
[0078] output = f{[X″ S X″ L ]W 1}W 2
[0079] In the above formula, X″ S , X″ L The one-dimensional transformation representation of the spatiotemporal features captured by the near-term and medium- and long-term channels, respectively, and the weight matrix dc is the concatenation vector dimension, d h is the hidden layer dimension, N is the number of predicted wind turbine objects, and f is the activation function.
[0080] In the embodiment of the present invention, the wind power prediction model is trained by using the Adam optimization algorithm, and the mean square error (MAE) is used as the loss function.
[0081] In the embodiment of the present invention, the test samples are input into the trained wind power prediction model to predict the power of wind turbines, and the mean absolute error (MAE), root mean square error (RMSE) and coefficient of determination (R 2 ) to evaluate the predictive performance of the algorithm.
[0082] Example 1
[0083] The data of a wind farm in northwest my country were obtained, including the historical power, historical wind speed and weather forecast wind speed data of each unit. The resolution of historical power and historical wind speed is 15 minutes, and the resolution of weather forecast wind speed is 1 hour. The nearest neighbor linear interpolation algorithm was used to interpolate missing values and outliers. Due to the differences in data dimensions and value ranges, in order to speed up model training and improve prediction results, the sampled data was standardized and the data values were limited to the interval [0, 1]. Then the sample data was divided into training samples and test samples according to a ratio of 9:1.
[0084] In this specific embodiment, the wind farm spatiotemporal data source is 33 unit objects A 1 ,…,A 33 , each unit has a data feature dimension of M = 3; the spatial data formed at any time t can be expressed as
[0085]
[0086] When the sequence length T = 64, the spatiotemporal data generated by 33 units can be expressed as a three-dimensional tensor To express.
[0087] X=(X 1 ,X 2 ,…,X 64 )
[0088] If we want to predict the state X of these N objects at the target time t+τ t+τ , considering the differences in steady-state and dynamic characteristics of spatiotemporal data at different time scales, and the fact that correlation usually weakens with the increase of time series length, it is necessary to conduct multi-scale joint learning. Without loss of generality, taking two time scales as an example, the spatiotemporal data X input to the prediction model is divided into two parts at different time scales: the near short-term observation sequence X Sand the medium- and long-term observation series X L , and other time scale situations can be expanded from it.
[0089]
[0090] In the above formula, d and c are the short-term sequence X S and the medium- and long-term series X L The length of the sequence is , and the relationship d+c=T is satisfied. The observation step length of the medium-to-near short-term sequence is 1, and the observation step length of the medium-to-long-term sequence is q. In this embodiment, d=32, c=32, q=12. Specifically, since the sampling unit of the data is 15 minutes, for short-term data, 32 observation data are sampled before time t with a step length of 1 basic unit (15 minutes), and for long-term data, 32 observation data are sampled before time t with a step length of 12 basic units (3 hours).
[0091] The wind power prediction model is established, which includes three modules: long-term feature learning module, short-term feature learning module and feature fusion output module.
[0092] (1) The long-term feature learning module uses a lightweight convolutional threshold unit to reduce the complexity of network training and improve the time series modeling capability while capturing the long-term trend characteristics of wind power data and the temporal and spatial correlation characteristics. Its working mechanism is as follows:
[0093]
[0094] In the above formula, * represents the convolution operation, represents the Hadamard product. t , Z t , H t and H′ t Represent the reset gate, update gate, current memory state and candidate memory state respectively, C t is the input three-dimensional tensor, W HZ , W CR , W HR and W HH is the corresponding two-dimensional convolution kernel, B Z , B R and B H is the corresponding bias, and σ is the Sigmoid activation function.
[0095] (2) The short-term feature learning module is based on a three-dimensional convolutional network, which is constructed by stacking two streamlined three-dimensional convolutional modules, Conv-1 (Conv3D-1, ReLU1) and Conv-2 (Conv3D-2, ReLU2). It can share the spatial information of related objects, learn the temporal correlation, and effectively extract the small-scale dynamic characteristics of adjacent short-term data.
[0096] (3) The feature fusion output module uses a fully connected network to fuse the features of (1) and (2) and finally outputs the wind power of each unit at the future moment.
[0097] The wind power prediction model is trained by the Adam optimization algorithm, and the mean square error (MAE) is used as the loss function. The learning rate is 0.001, the iteration is 40 times, and the network structure parameters of the three submodules are set as follows: the near-term short-term feature learning channel includes two three-dimensional convolution modules, the number of convolution kernels is 4 and 8, and the convolution kernel sizes are 6×1×6 and 3×1×3 respectively; the medium- and long-term feature learning channel has two layers of lightweight convolution threshold unit LCGRU units, the number of hidden layer units is 32, and the convolution kernel size is 1×1; the multi-channel feature fusion output module uses a fully connected network, sets one hidden layer, and the output layer is 33, the number of predicted objects.
[0098] The test samples are input into the trained model for wind power prediction, and the mean absolute error (MAE), root mean square error (RMSE) and coefficient of determination (R 2 )Evaluate the prediction performance of the algorithm. Error analysis is an important means to test whether the prediction model is effective. Both the mean absolute error and the root mean square error can evaluate the accuracy of the prediction results. The lower the better. The coefficient of determination is used to indicate the quality of the prediction model. The normal value range is [0,1]. The closer the result is to 1, the better the model fits the data and the stronger the explanatory power of the model.
[0099] Table 1 shows the prediction performance evaluation results of the embodiments of the present invention and the prior art. Specifically, the sample data time range is from January to December 2020, the data sampling resolution is 15 minutes, and the wind power data of the past 3 days is sampled at a step interval of 3 hours as the medium- and long-term data. The wind power data of 8 hours on the day and the weather forecast wind speed data are sampled at a step interval of 15 minutes as the near short-term data to predict the wind power of 33 units in the next 1 to 4 hours.
[0100] It can be seen from Table 1 that, under the prediction time of 1 to 4 hours, the RMSE of the present invention is reduced by about 6.8%, 11.4%, 1.0%, and 6.0% respectively compared with the prior art, and the MAE is also reduced by 7.5% to 24.0%; as the prediction time increases, the error indicators MAE and RMSE of each model increase accordingly. In addition, the coefficient of determination R 2 It is higher than the prior art. As the prediction time increases, the coefficient of determination decreases. This is because the shorter the prediction time, the better the fitting effect of the prediction model on the data, which is consistent with the theoretical time series characteristics. In summary, it shows that the prediction model of the present invention has a strong generalization ability and has certain advantages in practical engineering applications.
[0101] Table 1 Evaluation index results of the present invention and the prior art
[0102]
[0103] Those skilled in the art will appreciate that the embodiments of the present application may be provided as methods, systems, or computer program products. Therefore, the present application may adopt the form of a complete hardware embodiment, a complete software embodiment, or an embodiment in combination with software and hardware. Moreover, the present application may adopt the form of a computer program product implemented in one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) that include computer-usable program code.
[0104] The present application is described with reference to the flowcharts and / or block diagrams of the methods, devices (systems), and computer program products according to the embodiments of the present application. It should be understood that each process and / or box in the flowchart and / or block diagram, as well as the combination of the processes and / or boxes in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to generate a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowchart and / or block diagram. Figure 1 A process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.
[0105] These computer program instructions may also be stored in a computer-readable memory capable of directing a computer or other programmable data processing device to operate in a specific manner, so that the instructions stored in the computer-readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 A process or multiple processes and / or boxes Figure 1 A function specified in one or more boxes.
[0106] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operating steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing instructions for implementing the process in the computer or other programmable device. Figure 1 A process or multiple processes and / or boxes Figure 1 The steps for the functions specified in one or more boxes.
[0107] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the above embodiments, ordinary technicians in the relevant field should understand that the specific implementation methods of the present invention can still be modified or replaced by equivalents. Any modification or equivalent replacement that does not depart from the spirit and scope of the present invention should be covered within the scope of protection of the claims of the present invention.
Claims
1. A wind power prediction method integrating three-dimensional convolution and lightweight convolution threshold unit, It is characterized in that include: The spatiotemporal data of wind farm units are divided into short-term series X S and the medium- and long-term series X L , the wind farm unit spatiotemporal data includes: endogenous attribute data and external state data; the endogenous attribute data includes cut-in wind speed, cut-out wind speed and control strategy matrix; the external state data includes wind farm unit power, wind farm wind speed and meteorological forecast wind speed; the sequence is expressed as: Among them, the short-term series X S It means taking the current time point t as the benchmark, taking 1 sampling unit as the step length, and taking d data in sequence forward; the medium and long-term series X L It means taking the current time point t as the reference, taking q sampling units as the step length, and taking c data forward in sequence; X t represents the spatiotemporal data at time point t, Represents the j-dimensional spatiotemporal data of the i-th unit object at time t, N is the number of wind farm units, M is the spatiotemporal data dimension of wind farm units, T is the total number of data in the short-term sequence and the medium- and long-term sequence, satisfying T = d + c; A wind power prediction model integrating three-dimensional convolution and lightweight convolution threshold unit is established as follows: Two 3D convolutional modules are stacked to construct a 3D convolutional network to train the short-term sequence X S Capture spatiotemporal features for input learning; A double-layer lightweight convolution threshold unit is used to store medium- and long-term sequences X L To learn and capture spatiotemporal features for input; the double-layer lightweight convolution threshold unit is expressed as: Among them, * represents the convolution operation, represents the Hadamard product, R t , Z t , H t and H t 'represents the reset gate, update gate, current memory state and candidate memory state respectively, C t The input three-dimensional tensor is the medium- and long-term sequence X L , W HZ , W CR , W HR and W HH is the corresponding two-dimensional convolution kernel, B Z , B R and B H is the corresponding bias, σ is the Sigmoid activation function; A fully connected network is used to fuse the output spatiotemporal features, and finally a complete wind farm unit wind power prediction result is obtained; The wind power prediction model established by collecting historical spatiotemporal data of wind farm units for training; The real-time spatiotemporal data of wind farm units are input into the trained wind power prediction model to obtain the power prediction results of wind farm units.
2. According to claim 1, a wind power prediction method integrating three-dimensional convolution and lightweight convolution threshold unit, It is characterized in that Also includes: The nearest neighbor linear interpolation algorithm is used to interpolate missing values and outliers in the spatiotemporal data of wind farm units.
3. According to claim 2, a wind power prediction method integrating three-dimensional convolution and lightweight convolution threshold unit, It is characterized in that Also includes: The processed spatiotemporal data of wind farm units are standardized to limit the data values to the interval [0, 1].
4. According to claim 3, a wind power prediction method integrating three-dimensional convolution and lightweight convolution threshold unit, It is characterized in that The method of using a fully connected network to fuse the output spatiotemporal features includes: Flatten the captured spatiotemporal features into one-dimensional vectors and concatenate them. Use the fully connected layer to map the learned spatiotemporal features to the sample label space, and reconstruct the two spatiotemporal features into a complete feature map through the weight matrix.
5. According to claim 3, a wind power prediction method integrating three-dimensional convolution and lightweight convolution threshold unit, It is characterized in that The wind power prediction model is trained using the Adam optimization algorithm.
6. A wind power prediction method integrating three-dimensional convolution and lightweight convolution threshold unit according to claim 5, It is characterized in that During the training process, the mean square error is used as the loss function.
7. A wind power prediction system integrating three-dimensional convolution and lightweight convolution threshold unit, It is characterized in that A wind power prediction method for implementing the fusion three-dimensional convolution and lightweight convolution threshold unit according to any one of claims 1 to 6, the system comprising: A long-term feature learning module is used to construct a three-dimensional convolutional network to learn the long-term spatiotemporal data of wind farm units; the three-dimensional convolutional network is constructed by stacking two three-dimensional convolutional modules; The short-term feature learning module is used to construct a double-layer lightweight convolution threshold unit to learn the short-term spatiotemporal data of wind farm units; the double-layer lightweight convolution threshold unit is expressed as: Among them, * represents the convolution operation, represents the Hadamard product, R t , Z t , H t and H t 'represents the reset gate, update gate, current memory state and candidate memory state respectively, C t The input three-dimensional tensor is the medium- and long-term sequence X L , W HZ , W CR , W HR and W HH is the corresponding two-dimensional convolution kernel, B Z , B R and B H is the corresponding bias, σ is the Sigmoid activation function; The training module is used to collect the historical spatiotemporal data of wind farm units and divide the spatiotemporal data of wind farm units into short-term sequences X S and the medium- and long-term series X L , the wind farm unit spatiotemporal data includes: endogenous attribute data and external state data; the endogenous attribute data includes cut-in wind speed, cut-out wind speed and control strategy matrix; the external state data includes wind farm unit power, wind farm wind speed and meteorological forecast wind speed; the sequence is expressed as: Among them, the short-term series X S It means taking the current time point t as the benchmark, taking 1 sampling unit as the step length, and taking d data in sequence forward; the medium and long-term series X L It means taking the current time point t as the reference, taking q sampling units as the step length, and taking c data forward in sequence; X t represents the spatiotemporal data at time point t, Represents the j-dimensional spatiotemporal data of the i-th unit object at time t, N is the number of wind farm units, M is the spatiotemporal data dimension of wind farm units, T is the total number of data in the short-term sequence and the medium- and long-term sequence, satisfying T = d + c; Take the short-term series X S Capture spatiotemporal features for learning input; Taking the medium and long term series X L Capture spatiotemporal features for learning input; A fully connected network is used to fuse the output spatiotemporal features, and finally a complete wind farm unit wind power prediction result is obtained; The output module is used to input the real-time spatiotemporal data of wind farm units into the trained wind power prediction model to obtain the power prediction results of the wind farm units.
