Wind power prediction method and system based on multivariate time sequence full-connection space-time diagram

Abstract

The invention provides a wind power prediction method and system based on a multivariate time sequence full-connection space-time diagram, and relates to the technical field of wind power prediction. Various sensor data of wind power, wind speed, temperature, humidity, wind direction, yaw angle and pitch angle of a wind power plant at different timestamps are collected to form a multivariate time sequence data set, and support is provided for more accurate wind power prediction; the dot product of each sensor is calculated to construct a full connection graph model, a complex time-space relationship is dynamically captured, the understanding ability of the model on the dependence between the sensors is effectively improved, and the prediction accuracy is enhanced; according to the wind power prediction model, the calculation complexity is effectively reduced while the prediction precision is ensured.

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 based on a multivariate time series fully connected space-time graph. Background Art

[0002] Wind power generation is an important clean energy source, but the uncertainty of wind speed leads to the instability of wind turbine output power, which in turn makes wind power generation present randomness and volatility. With the continuous increase in the number of wind farms and installed capacity, the randomness and volatility of wind power generation after wind power is connected to the grid will bring huge challenges to the safe and stable operation of the power grid.

[0003] Therefore, accurate prediction of wind power is of great significance to ensure the stability and security of the power system during large-scale wind power grid connection. Existing wind power prediction methods have a single data source and mostly rely on numerical weather forecasts as input data. However, the accuracy of numerical weather forecasts is easily affected by initial conditions and boundary conditions, especially under complex terrain and changeable meteorological conditions. This dependence will further amplify the prediction error, resulting in reduced accuracy of wind power prediction results.

[0004] Existing wind power prediction methods mainly rely on physical models and statistical models. However, physical models rely heavily on meteorological data and have difficulty dealing with complex nonlinear relationships. Although statistical models can capture the changing patterns of wind power to a certain extent, they have limited performance when dealing with high-dimensional and multi-variable variables, and it is difficult to accurately predict the spatiotemporal correlation of wind power.

[0005] In addition, existing wind power prediction models have high computational complexity, low efficiency, and lack flexibility in parameter setting. Summary of the invention

[0006] In view of the deficiencies in the prior art, the present invention provides a wind power prediction method and system based on a multivariate time-series fully connected spatiotemporal graph, which solves the problem of low accuracy in wind power prediction.

[0007] To achieve the above objectives, the present invention is implemented through the following technical solutions:

[0008] A wind power prediction method based on a multivariate time series fully connected space-time graph, the prediction method comprising the following steps:

[0009] S1. Collect various sensor data of wind power, wind speed, temperature, humidity, wind direction, yaw angle and pitch angle at different time stamps of the wind farm to form a multivariate time series data sample set, and obtain a standard data sample set after preprocessing;

[0010] S2, divide the standard data sample set into multiple small blocks according to the timestamp, realize the comprehensive interconnection between sensors through dot product calculation, build a fully connected graph, and introduce the attenuation matrix considering the influence of time distance;

[0011] S3. Design sliding windows and apply graph neural networks in each window to capture the spatiotemporal dependencies within the window. Use message passing neural networks to process the fully connected graph, learn high-level features through time pooling operations, and output wind power prediction values.

[0012] Preferably, in S1, the preprocessing comprises the following steps:

[0013] The missing values ​​are used to replace the outliers in the data sample set, and then the neighboring interpolation processing is performed at the missing values. Then, the standardization processing is performed to obtain a standard data sample set.

[0014] Preferably, in S2, segmenting the data includes the following steps:

[0015] Each small block is represented as:

[0016]

[0017] in, Indicates the number of blocks to be split.

[0018] f represents the size of the block;

[0019] t represents the block index of the timestamp;

[0020] The data after segmentation by timestamp is

[0021] Among them, x t,i represents the data record of the i-th sensor at timestamp t;

[0022] Use encoder f c (·|W c ) Process the data in each block, learn each sensor-level feature, and integrate the position information into the sensor feature through position encoding;

[0023] Among them, [·] indicates truncation operation;

[0024] W c Represents a set of learnable parameters.

[0025] Preferably, in S2, the fully connected graph construction comprises the following steps:

[0026] By calculating the dot product between the sensors of wind power, wind speed, temperature, humidity, wind direction, yaw angle and pitch angle, a fully connected graph is constructed to quantify the similarity between the sensors, which is defined as:

[0027]

[0028] Among them, g s (z t,i ) represents a transformation function, which is used to enhance the expressive power of the input vector z;

[0029] z t,i represents the feature representation of the i-th sensor at time t;

[0030] z r,j represents the feature representation of the jth sensor at time r;

[0031] (g s (z r,j )) T represents transpose;

[0032] Using a learnable weight matrix g s (z)=zW s Enhance expressive ability;

[0033] Among them, W s are learnable weights;

[0034] And the correlation is limited to [0, 1] through the softmax function. Finally, the fully connected graph is obtained:

[0035] G = (Z, E);

[0036] in,

[0037] Z represents the data set;

[0038] z t,i Represents snippets of data collected from timestamps and sensors;

[0039] E represents the adjacency matrix of the fully connected graph;

[0040] e tr,ij Represents the correlation between sensors between time bins.

[0041] Preferably, in S2, the attenuation matrix is ​​expressed as:

[0042]

[0043] Among them, each element c tr,ij =δ t-r , to calculate the matrix used to adjust the sensor correlation between blocks, ensure that the temporally adjacent sensors have a strong correlation, and improve the modeling effect of spatiotemporal dependency;

[0044] N represents the number of sensors;

[0045] Indicates the number of split blocks;

[0046] tr represents the time index;

[0047] δ represents the decay rate;

[0048] c tr,ij represents the element in the decay matrix, which represents the correlation between sensors i and j between time bins t and r;

[0049] tr represents the time difference.

[0050] Preferably, in said S3, the message passing neural network includes: a propagation phase and an update phase;

[0051] In the propagation phase, the information of neighboring nodes is transmitted to the central node. Given the central node of the wth window in the lth layer, It has a set of adjacent nodes in the M blocks in the same window The relationship between the central node and its neighbors is:

[0052]

[0053] The characteristics after propagation are:

[0054]

[0055] Where M represents the number of blocks in each window;

[0056] l represents the number of layers or levels of the neural network;

[0057] w represents the index of the center node of the window;

[0058] represents the feature representation of the jth sensor in the wth window in the lth layer and at r time;

[0059] Represents the correlation between the central node and the neighboring nodes in the lth layer;

[0060] Represents the feature representation of the central node after l layers of propagation;

[0061] In the updating stage, nonlinear functions are used to update the propagated features to improve the expressiveness of the features.

[0062] Preferably, in S3, the sensor features after propagation and updating are then subjected to a time pooling operation to extract high-level features of each window; and the average pooling strategy is used to integrate multiple blocks of data in the window to obtain the sensor features of the window w:

[0063]

[0064] After stacking the features of all windows, an advanced fully connected graph convolutional network is constructed as the output layer, and the predicted wind power is finally output. The mean absolute error and root mean square error are calculated based on the predicted wind power as the evaluation indicators of the final wind power prediction effect.

[0065] A wind power prediction system based on a multivariate time series fully connected spatiotemporal graph, the prediction system comprising: a data collection module, a fully connected graph construction module, a moving window module, a message passing neural network module and an output module;

[0066] The data collection module is used to collect multiple sensor data of wind power, wind speed, temperature, humidity, wind direction, yaw angle and pitch angle at different time stamps of the wind farm to form a multivariate time series data sample set, and obtain a standard data sample set after preprocessing;

[0067] The fully connected graph building module is used to divide the standard data sample set into multiple small blocks according to the timestamp, realize the comprehensive interconnection between sensors through dot product calculation, build a fully connected graph, and introduce the attenuation matrix considering the influence of time distance;

[0068] The moving window module is used to design a sliding window, and a graph neural network is applied in each window to capture the spatiotemporal dependencies within the window;

[0069] The message passing neural network module is used to process the fully connected graph using a message passing neural network;

[0070] The output module is used to learn high-level features through time pooling operations and output wind power prediction values.

[0071] The present invention provides a wind power prediction method and system based on a multivariate time series fully connected space-time graph. Compared with the prior art, it has the following beneficial effects:

[0072] In the present invention, a variety of sensor data of wind power, wind speed, temperature, humidity, wind direction, yaw angle and pitch angle at different time stamps of the wind farm are collected to form a multivariate time series data set, providing support for more accurate wind power prediction; the dot product of each sensor is calculated to construct a fully connected graph model, dynamically capture complex spatiotemporal relationships, effectively improve the model's ability to understand the dependencies between sensors, and enhance the accuracy of prediction; the wind power prediction model effectively reduces the computational complexity while ensuring prediction accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0073] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0074] Figure 1 is a flow chart of a prediction method in an embodiment of the present invention;

[0075] Figure 2 4 is a structural diagram of a prediction system in an embodiment of the present invention. DETAILED DESCRIPTION

[0076] In order to make the purpose, technical solutions and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention are clearly and completely described. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0077] The embodiments of the present application solve the problem of low accuracy of wind power prediction by providing a wind power prediction method and system based on a multivariate time-series fully connected spatiotemporal graph.

[0078] In order to better understand the above technical solution, the above technical solution will be described in detail below in conjunction with the accompanying drawings and specific implementation methods.

[0079] Example:

[0080] like Figure 1 As shown, the present invention provides a wind power prediction method based on a multivariate time series fully connected space-time graph, and the prediction method includes the following steps:

[0081] S1. Data collection and preprocessing

[0082] Collect various sensor data of wind power, wind speed, temperature, humidity, wind direction, yaw angle and pitch angle at different time stamps of the wind farm to form a multivariate time series data sample set. Use missing values ​​to replace abnormal values ​​in the data sample set, perform neighbor interpolation processing at the missing values, and then perform standardization processing to obtain a standard data sample set.

[0083] S2. Build a fully connected graph

[0084] (1) Data segmentation

[0085] The standard data sample set processed in the first step is used as input data and divided into multiple small blocks according to the timestamp. Each small block contains data on wind power, wind speed, temperature, humidity, wind direction, yaw angle and pitch angle. Each small block is represented as:

[0086]

[0087] in, Indicates the number of blocks to be split.

[0088] f represents the size of the block;

[0089] t represents the block index of the timestamp;

[0090] The data after segmentation by timestamp is

[0091] Among them, x t,i represents the data record of the i-th sensor at timestamp t;

[0092] Then, using the encoder f c (·|W c ) Process the data in each block, learn each sensor-level feature, and integrate the position information into the sensor feature through position encoding;

[0093] Among them, [·] indicates truncation operation;

[0094] W c Represents a set of learnable parameters.

[0095] (2) Fully connected graph construction

[0096] By calculating the dot product between the sensors of wind power, wind speed, temperature, humidity, wind direction, yaw angle and pitch angle, a fully connected graph is constructed to quantify the similarity between the sensors, which is defined as:

[0097]

[0098] Among them, g s (z t,i ) represents a transformation function, which is used to enhance the expressive power of the input vector z;

[0099] z t,i represents the feature representation of the i-th sensor at time t;

[0100] z r,j represents the feature representation of the jth sensor at time r;

[0101] (g s (z r,j )) TIndicates transpose.

[0102] Then, using the learnable weight matrix g s (z)=zW s Enhance expressive ability;

[0103] Among them, W s are learnable weights;

[0104] And the correlation is limited to [0, 1] through the softmax function. Finally, the fully connected graph is obtained:

[0105]

[0106] Where Z represents the data set;

[0107] z t,i Represents snippets of data collected from timestamps and sensors;

[0108] E represents the adjacency matrix of the fully connected graph;

[0109] e tr,jj represents the correlation between sensors between time blocks;

[0110] The fully connected graph G not only contains the temporal correlation between timestamps and the spatial correlation in each timestamp, but also captures the cross-temporal and spatial dependencies between wind power, wind speed, temperature, humidity, wind direction, yaw angle, and pitch angle sensors.

[0111] (3) Attenuation Matrix

[0112] Since the correlation between wind power, wind speed, temperature, humidity, wind direction, yaw angle and pitch angle gradually decays as the time interval increases, the decay matrix is ​​defined by setting the decay rate δ:

[0113]

[0114] Among them, each element c tr,ij =δ t-r , to calculate the matrix used to adjust the sensor correlation between blocks, ensure that the temporally adjacent sensors have a strong correlation, and improve the modeling effect of spatiotemporal dependency;

[0115] N represents the number of sensors;

[0116] Indicates the number of split blocks;

[0117] tr represents the time index;

[0118] δ represents the decay rate;

[0119] c tr,ijrepresents the element in the decay matrix, which represents the correlation between sensors i and j between time bins t and r;

[0120] tr represents the time difference.

[0121] S3, Fully Connected Graph Convolutional Network

[0122] (1) Design sliding window

[0123] A moving window of a specific size M slides with a stride s, and a graph neural network is applied within each window to capture the spatiotemporal dependencies within the window.

[0124] (2) Message Passing Neural Network

[0125] In each window, a message passing neural network is used to process the fully connected graph. It includes two stages: propagation and update. In the propagation stage, the information of neighboring nodes is passed to the central node. Given the central node of the wth window in the lth layer, It has a set of adjacent nodes in the M blocks in the same window The relationship between the central node and its neighbors is The characteristics after propagation are In the updating stage, nonlinear functions are used to update the propagated features to improve the expressiveness of the features.

[0126] Where M represents the number of blocks in each window;

[0127] l represents the number of layers or levels of the neural network;

[0128] w represents the index of the center node of the window;

[0129] represents the feature representation of the jth sensor in the wth window in the lth layer and at r time;

[0130] Represents the correlation between the central node and the neighboring nodes in the lth layer;

[0131] Represents the feature representation of the central node after l layers of propagation.

[0132] (3) Average Pooling

[0133] After propagation and updating of sensor features, high-level features of each window are extracted through temporal pooling operation. The average pooling strategy is used to integrate multiple blocks of data in the window to obtain the sensor features of window w:

[0134]

[0135] After stacking the features of all windows, an advanced fully connected graph convolutional network is constructed as the output layer, and the predicted wind power is finally output. The mean absolute error and root mean square error are calculated based on the predicted wind power as the evaluation indicators of the final wind power prediction effect.

[0136] A wind power prediction system based on a multivariate time series fully connected spatiotemporal graph, the prediction system comprising: a data collection module, a fully connected graph construction module, a moving window module, a message passing neural network module and an output module;

[0137] The data collection module is used to collect multiple sensor data of wind power, wind speed, temperature, humidity, wind direction, yaw angle and pitch angle at different time stamps of the wind farm to form a multivariate time series data sample set, and obtain a standard data sample set after preprocessing;

[0138] The fully connected graph building module is used to divide the standard data sample set into multiple small blocks according to the timestamp, realize the comprehensive interconnection between sensors through dot product calculation, build a fully connected graph, and introduce the attenuation matrix considering the influence of time distance;

[0139] The moving window module is used to design a sliding window, and a graph neural network is applied in each window to capture the spatiotemporal dependencies within the window;

[0140] The message passing neural network module is used to process the fully connected graph using a message passing neural network;

[0141] The output module is used to learn high-level features through time pooling operations and output wind power prediction values.

[0142] In summary, compared with the prior art, the present invention has the following beneficial effects:

[0143] In the embodiment of the present invention, a variety of sensor data of wind power, wind speed, temperature, humidity, wind direction, yaw angle and pitch angle at different time stamps of the wind farm are collected to form a multivariate time series data set, providing support for more accurate wind power prediction; the dot product of each sensor is calculated to construct a fully connected graph model, dynamically capture complex spatiotemporal relationships, effectively improve the model's ability to understand the dependencies between sensors, and enhance the accuracy of prediction; the wind power prediction model effectively reduces the computational complexity while ensuring prediction accuracy.

[0144] It should be noted that, in this article, relational terms such as first and second, etc. are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Moreover, the terms "include", "comprise" or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or device. In the absence of further restrictions, the elements defined by the sentence "comprise a ..." do not exclude the existence of other identical elements in the process, method, article or device including the elements.

[0145] The above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit the same. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that the technical solutions described in the aforementioned embodiments may still be modified, or some of the technical features may be replaced by equivalents. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A wind power prediction method based on a multivariate time series fully connected spatiotemporal graph, characterized in that: The prediction method comprises the following steps: S1. Collect various sensor data of wind power, wind speed, temperature, humidity, wind direction, yaw angle and pitch angle at different time stamps of the wind farm to form a multivariate time series data sample set, and obtain a standard data sample set after preprocessing; S2, divide the standard data sample set into multiple small blocks according to the timestamp, realize the comprehensive interconnection between sensors through dot product calculation, build a fully connected graph, and introduce the attenuation matrix considering the influence of time distance; S3. Design sliding windows and apply graph neural networks in each window to capture the spatiotemporal dependencies within the window. Use message passing neural networks to process the fully connected graph, learn high-level features through time pooling operations, and output wind power prediction values.

2. The method for wind power prediction based on a multivariate time series fully connected spatiotemporal graph according to claim 1, characterized in that: In S1, the preprocessing includes the following steps: The missing values ​​are used to replace the outliers in the data sample set, and then the neighboring interpolation processing is performed at the missing values. Then, the standardization processing is performed to obtain a standard data sample set.

3. The method for wind power prediction based on a multivariate time series fully connected spatiotemporal graph according to claim 1, characterized in that: In S2, segmenting the data includes the following steps: Each small block is represented as: in, Indicates the number of blocks to be split. f represents the size of the block; t represents the block index of the timestamp; The data after segmentation by timestamp is Among them, x t,i represents the data record of the i-th sensor at timestamp t; Use encoder f c (·|W c ) Process the data in each block, learn each sensor-level feature, and integrate the position information into the sensor feature through position encoding; Among them, [·] indicates truncation operation; W c Represents a set of learnable parameters.

4. The method for wind power prediction based on a multivariate time series fully connected spatiotemporal graph according to claim 1, characterized in that: In S2, the fully connected graph construction includes the following steps: By calculating the dot product between the sensors of wind power, wind speed, temperature, humidity, wind direction, yaw angle and pitch angle, a fully connected graph is constructed to quantify the similarity between the sensors, which is defined as: Among them, g s (z t,i ) represents a transformation function, which is used to enhance the expressive power of the input vector z; z t,i represents the feature representation of the i-th sensor at time t; z r,j represents the feature representation of the jth sensor at time r; (g s (z r,j )) T represents transpose; Using a learnable weight matrix g s (z)=zW s Enhance expressive ability; Among them, W s are learnable weights; And the correlation is limited to [0, 1] through the softmax function. Finally, the fully connected graph is obtained: G = (Z, E); Z represents the data set; z t,i Represents snippets of data collected from timestamps and sensors; E represents the adjacency matrix of the fully connected graph; e tr,ij Represents the correlation between sensors between time bins.

5. The method for wind power prediction based on multivariate time series fully connected spatiotemporal graph according to claim 1, characterized in that: In S2, the attenuation matrix is ​​expressed as: Among them, each element c tr,ij =δ t-r , to calculate the matrix used to adjust the sensor correlation between blocks, ensure that the temporally adjacent sensors have a strong correlation, and improve the modeling effect of spatiotemporal dependency; N represents the number of sensors; Indicates the number of split blocks; tr represents the time index; δ represents the decay rate; c tr,ij represents the element in the decay matrix, which represents the correlation between sensors i and j between time bins t and r; tr represents the time difference.

6. The method for wind power prediction based on multivariate time series fully connected spatiotemporal graph according to claim 1, characterized in that: In S3, the message passing neural network includes: a propagation phase and an update phase; In the propagation phase, the information of neighboring nodes is transmitted to the central node. Given the central node of the wth window in the lth layer, It has a set of adjacent nodes in the M blocks in the same window The relationship between the central node and its neighbors is: The characteristics after propagation are: Where M represents the number of blocks in each window; l represents the number of layers or levels of the neural network; w represents the index of the center node of the window; represents the feature representation of the jth sensor in the wth window in the lth layer and at r time; Represents the correlation between the central node and the neighboring nodes in the lth layer; Represents the feature representation of the central node after l layers of propagation; In the updating stage, nonlinear functions are used to update the propagated features to improve the expressiveness of the features.

7. The method for wind power prediction based on multivariate time series fully connected spatiotemporal graph according to claim 1, characterized in that: In S3, after propagation and updating, the sensor features are used to extract high-level features of each window through a time pooling operation; the average pooling strategy is used to integrate multiple blocks of data in the window to obtain the sensor features of window w: After stacking the features of all windows, an advanced fully connected graph convolutional network is constructed as the output layer, and the predicted wind power is finally output. The mean absolute error and root mean square error are calculated based on the predicted wind power as the evaluation indicators of the final wind power prediction effect.

8. A wind power prediction system based on a multivariate time series fully connected spatiotemporal graph, characterized in that: The prediction system includes: a data collection module, a fully connected graph construction module, a moving window module, a message passing neural network module and an output module; The data collection module is used to collect multiple sensor data of wind power, wind speed, temperature, humidity, wind direction, yaw angle and pitch angle at different time stamps of the wind farm to form a multivariate time series data sample set, and obtain a standard data sample set after preprocessing; The fully connected graph building module is used to divide the standard data sample set into multiple small blocks according to the timestamp, realize the comprehensive interconnection between sensors through dot product calculation, build a fully connected graph, and introduce the attenuation matrix considering the influence of time distance; The moving window module is used to design a sliding window, and a graph neural network is applied in each window to capture the spatiotemporal dependencies within the window; The message passing neural network module is used to process the fully connected graph using a message passing neural network; The output module is used to learn high-level features through time pooling operations and output wind power prediction values.

Images