A method and device for predicting offshore wind farm power by integrating temporal and spatial information

Abstract

The present invention provides a method and device for predicting offshore wind farm power by integrating spatiotemporal information, the method comprising: obtaining mixed time series data of all wind turbines in an offshore wind farm to construct a tensor data set, dividing the tensor data set into a training set, a validation set, and a test set; using the training set to train the TSMixer model to obtain an offshore wind farm power prediction model; using the validation set to adjust and verify the hyperparameters of the offshore wind farm power prediction model; inputting the test set into the offshore wind farm power prediction model to achieve offshore wind farm power prediction. The present invention adds a spatial feature attention layer and feature sequence mixers of multiple different time scales, and considers the coupling characteristics between wind turbines at multiple different locations in the same wind farm and the feature information of the historical time series data of a single wind turbine at different time scales in the offshore wind power prediction task, thereby improving the feature information extraction capability of the model.

Description

Technical Field

[0001] The present invention relates to the technical field of offshore wind farm power prediction, and in particular to an offshore wind farm power prediction method and device integrating time and space information. Background Art

[0002] Compared with onshore wind power, offshore wind power faces a more complex and changeable meteorological environment, which is more difficult to predict, and the safe and reliable operation of wind farms faces greater challenges. In addition, the specific heat capacity of seawater is higher than that of land, and the thermal effect and wake effect of wind flow are significant. The common wind power physical model is difficult to adapt to the reality of rapid evolution of offshore meteorological conditions and complex and changeable sea conditions. Therefore, it is necessary to carry out research on power prediction methods more suitable for offshore wind power.

[0003] The existing wind power prediction methods can be roughly divided into physical model methods, time series model methods, machine learning methods, and deep learning methods. Due to the changeable meteorological conditions at sea, the output power of offshore wind power is deeply coupled with the sea conditions, which makes it difficult for general physical model methods and time series model methods to be applied to offshore wind power prediction. In addition, there is a complex nonlinear relationship between offshore wind power output and meteorological conditions, and machine learning methods are also difficult to accurately predict the output of offshore wind power. In recent years, deep learning methods have developed rapidly, and deep networks such as LSTM and GRU have been proposed and continuously improved to suit different time series prediction tasks. These deep learning methods all require model training through a structure similar to a recurrent neural network, which will consume more computing efficiency. For this reason, Google launched the TSMixer model in 2023, which not only uses a linear structure to improve computing efficiency, but also emphasizes the extraction of information between multi-variable inputs, thereby improving the accuracy of time series prediction.

[0004] However, the existing deep learning prediction methods for wind power still lack the ability to extract information about the spatial characteristics of different spatial positions between multiple wind turbines in an offshore wind farm, as well as the ability to extract information at different scales for the same multivariate time series. Therefore, how to better extract feature information of different dimensions, design a deep model that is more suitable for offshore wind power prediction, and realize the development of offshore wind power prediction methods and devices are issues that people in this field need to solve urgently.

[0005] In summary, there is an urgent need for an offshore wind farm power prediction method and device that integrates spatiotemporal information to solve the problems existing in the prior art. Summary of the invention

[0006] The purpose of the present invention is to provide an offshore wind farm power prediction method integrating spatiotemporal information, aiming to solve the problem of accurately predicting the short-term power of offshore wind farms and facilitate the smooth implementation of power system planning. The specific technical solution is as follows:

[0007] A method for predicting offshore wind farm power by integrating spatiotemporal information, comprising:

[0008] Obtain mixed time series data of all wind turbines in the offshore wind farm to construct a tensor dataset, and divide the tensor dataset into a training set, a validation set, and a test set;

[0009] The TSMixer model is trained using the training set to obtain an offshore wind farm power prediction model;

[0010] The validation set is used to adjust and verify the hyperparameters of the offshore wind farm power prediction model;

[0011] Input the test set into the offshore wind farm power prediction model to realize offshore wind farm power prediction;

[0012] The TSMixer model includes:

[0013] The time series input layer is used to output the input mixed time series data as a mixed time series feature map Y ;in, , N is the total number of wind turbines in the offshore wind farm, T is the length of the mixed time series, C is the number of features of the mixed time series, represents Euclidean space;

[0014] The mixer layer includes a time series mixer and a parallel setup m A feature sequence mixer of different time scales, wherein the MLP module in the time series mixer is cascaded with a spatial feature attention layer; the mixed time series feature map Y Input to the mixer layer and output is m Feature maps at different time scales Y M , m is a natural number greater than or equal to 2;

[0015] Time domain projection layer, m Feature maps at different time scales Y M After training separately, reconstruct them into prediction time series with the same specified output sequence length;

[0016] Output prediction layer, through the adder m The predicted time series of the specified output sequence length are added together to obtain the final predicted time series for output.

[0017] In the above technical solution, the mixed time series data includes the spatial position data of the wind turbine group. and multi-dimensional time series data; the multi-dimensional time series data includes time series data of wind speed, wind direction, air density, temperature, humidity, turbulence intensity and wind turbine power; wherein, For the The horizontal axis of the wind turbine is For the The vertical coordinate of a wind turbine is For the The blade height of a wind turbine.

[0018] Preferably in the above technical scheme, the spatial position data of the wind turbine is obtained by: obtaining the distribution positions of all wind turbines in the offshore wind farm, plotting the positions of all wind turbines in the offshore wind farm on a two-dimensional coordinate system by grid division, performing two-dimensional position encoding on each grid, recording the horizontal and vertical coordinates and blade height of the wind turbine in the grid, and obtaining the spatial position data of the wind turbine.

[0019] In the above technical solution, the mixed time series feature map is preferably Y The processing flow in the mixer layer is:

[0020] The time series mixer receives the mixed time series feature graph Y After that, a two-dimensional batch normalization operation and a transposition operation are performed to obtain a mixed time series feature map. Y T , mixed time series feature map Y T After feature extraction by the MLP module, a higher-dimensional mixed time series feature map is obtained. Y' , mixed time series feature map Y' After the spatial feature attention layer processing and transposition operation, a high-dimensional mixed time series feature map is obtained. Y H , the high-dimensional mixed time series feature map Y H Mixed time series feature plot Y After residual connection, a new high-dimensional mixed time series feature map is obtained Y HL ;

[0021] New high-dimensional mixed time series feature map Y HL After two-dimensional batch normalization operation, input them into m The uniform sampling operation of different time scales is performed in the feature sequence mixer of different time scales, and then input into their respective MLP modules for training, and finally output mFeature maps at different time scales Y HH ;

[0022] High-dimensional mixed time series feature map Y HL Respectively m Feature maps at different time scales Y HH After the residual connection, the output of the mixer layer is obtained m Feature maps at different time scales Y M .

[0023] In the above technical solution, the spatial feature attention layer first performs an operation on the mixed time series feature map. Y' Perform sampling operations to generate mixed time series feature maps Y 1, , the mixed time series feature map Y 1 is input to the mean pooling layer on the first dimension, and then activated by the sigmoid function after training through the 1×1 convolution layer, the nonlinear pooling layer and the 1×1 convolution layer. Finally, the output layer is reconstructed to ensure that the shape and size of the output feature map are consistent with the mixed time series feature map of the input mixer layer. Y consistent; among them K The number of wind turbines sampled, K <N / 4.

[0024] In the above technical solution, preferably, the spatial feature attention layer is used for the mixed time series feature map Y' The sampling operation is as follows: taking the location of a single wind turbine as the center, sampling the K Mixed time series data of wind turbines to generate the sampled mixed time series feature graph Y 1, .

[0025] In the above technical solution, preferably, each feature sequence mixer outputs a feature map Y M Input into the corresponding time domain projection layer respectively, each feature map Y M The transposition operation is first performed in the corresponding time domain projection layer, and then reconstructed into a predicted time series with the same specified output sequence length after training through the respective fully connected layers.

[0026] Preferably in the above technical scheme, the MLP module in the time series mixer is composed of a fully connected layer, a ReLU activation function and a dropout layer; the MLP module in the feature sequence mixer is composed of a linear connection of five layers of modules: a fully connected layer, a ReLU activation function, a dropout layer, a fully connected layer and a dropout layer.

[0027] Preferably, in the above technical solution, the tensor data set is randomly divided into a training set, a validation set and a test set in proportion.

[0028] The present invention also provides an offshore wind farm power prediction device that integrates time and space information. The prediction device includes a processor and a memory. The memory stores the offshore wind farm power prediction method that integrates time and space information. The processor is used to execute the offshore wind farm power prediction method that integrates time and space information.

[0029] The application of the technical solution of the present invention has the following beneficial effects:

[0030] The offshore wind farm power prediction method of the present invention integrates multiple time series data and wind turbine location information data, has stronger reliability, and adopts a novel deep time series network TSMixer as a skeleton model, which can more accurately predict the offshore wind farm power.

[0031] The offshore wind farm power prediction method of the present invention improves the deep time series network TSMixer model, adds a spatial feature attention layer and a plurality of feature sequence mixers at different time scales, and considers the coupling characteristics between wind turbines at multiple locations in the same wind farm and the feature information of the historical time series data of a single wind turbine at different time scales in the offshore wind power prediction task, thereby improving the feature information extraction capability of the model.

[0032] In addition to the above-described purposes, features and advantages, the present invention has other purposes, features and advantages. The present invention will be further described in detail with reference to the accompanying drawings. BRIEF DESCRIPTION OF THE DRAWINGS

[0033] The drawings constituting a part of this application are used to provide a further understanding of the present invention. The exemplary embodiments of the present invention and their descriptions are used to explain the present invention and do not constitute an improper limitation of the present invention. In the drawings:

[0034] Figure 1 is a flow chart of an offshore wind farm power prediction method integrating spatiotemporal information according to the present invention;

[0035] Figure 2 is a schematic diagram of the TSMixer model in the present invention;

[0036] Figure 3 yes Figure 2 Schematic diagram of the middle mixer layer;

[0037] Figure 4 yes Figure 3 Schematic diagram of the spatial feature attention layer.

[0038] Figure 5 yes Figure 2 Schematic diagram of the time domain projection layer. DETAILED DESCRIPTION

[0039] In order to facilitate the understanding of the present invention, the present invention will be described more fully below, and preferred embodiments of the present invention are provided. However, the present invention can be implemented in many different forms and is not limited to the embodiments described herein. On the contrary, the purpose of providing these embodiments is to make the understanding of the disclosure of the present invention more thorough and comprehensive.

[0040] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as those commonly understood by those skilled in the art of the present invention. The terms used in the specification of the present invention herein are only for the purpose of describing specific embodiments and are not intended to limit the present invention.

[0041] Example:

[0042] like Figure 1 As shown, this embodiment provides an offshore wind farm power prediction method integrating spatiotemporal information, including the following processing steps:

[0043] S1, obtaining mixed time series data of all wind turbines in the offshore wind farm to construct a tensor dataset, and dividing the tensor dataset into a training set, a validation set, and a test set;

[0044] S2. Use the training set to train the TSMixer model to obtain an offshore wind farm power prediction model;

[0045] S3. Use the validation set to adjust and verify the hyperparameters of the offshore wind farm power prediction model;

[0046] S4. Input the test set into the offshore wind farm power prediction model to realize the offshore wind farm power prediction.

[0047] Preferably, the mixed time series data includes spatial position data of the wind turbine and multi-dimensional time series data, the multi-dimensional time series data including time series data of wind speed, wind direction, air density, temperature, humidity, turbulence intensity and wind turbine power; wherein, For the The horizontal axis of the wind turbine is For the The vertical coordinate of a wind turbine is For the The blade height of a wind turbine.

[0048] Furthermore, the spatial position data of the wind turbine is obtained by obtaining the distribution positions of all wind turbines in the offshore wind farm, plotting the positions of all wind turbines in the offshore wind farm on a two-dimensional coordinate system by grid division, performing two-dimensional position encoding on each grid, recording the horizontal and vertical coordinates and blade height of the wind turbine in the grid, and obtaining the spatial position data of the wind turbine.

[0049] Preferably, the grid has a side length of n Meter square, where n The value range of is [1, +∞). In this embodiment n Take 10 meters.

[0050] Furthermore, in order to facilitate subsequent data processing and calculation, the positions of all wind turbines in the offshore wind farm are plotted in the same quadrant of the two-dimensional coordinate system during grid division, preferably the first quadrant.

[0051] Preferably, in this embodiment, the tensor data set is randomly divided into a training set, a validation set, and a test set in a ratio of 8:1:1.

[0052] Specifically, those skilled in the art know that the TSMixer model (Time Series Mixer, TSMixer) is a new full MLP architecture released by Google, which is specifically used for time series prediction. Different from the existing TSMixer model, the TSMixer model is optimized and improved in this embodiment to achieve more accurate prediction of offshore wind farm power.

[0053] Figure 2 Schematic diagram of the TSMixer model in this embodiment, the TSMixer model includes:

[0054] The time series input layer is used to output the input mixed time series data as a mixed time series feature map Y ,in, , N is the total number of wind turbines in the offshore wind farm, T is the length of the mixed time series, C is the number of features of the mixed time series, represents Euclidean space;

[0055] Preferably, the mixed time series data of this embodiment includes the spatial position data of the wind turbine group And multi-dimensional time series data, realizing the fusion of spatial feature dimension and multi-scale time feature dimension. In this embodiment, the mixed time series includes 10 features: wind speed, wind direction, air density, temperature, humidity, turbulence intensity, wind turbine power, wind turbine horizontal coordinate, wind turbine vertical coordinate and wind turbine blade height.

[0056] The mixer layer includes a time series mixer and a parallel setup m A feature sequence mixer with different time scales, in which the MLP module in the time series mixer is cascaded with a spatial feature attention layer, such as Figure 3 shown.

[0057] Specifically, the mixed time series feature map Y The processing flow in the mixer layer is:

[0058] The time series mixer receives the mixed time series feature graph Y After that, a two-dimensional batch normalization operation and a transposition operation are performed to obtain a mixed time series feature map. Y T , and then after feature extraction through the MLP module, a higher-dimensional mixed time series feature map is obtained Y' , mixed time series feature map Y' After the spatial feature attention layer processing and transposition operation, a high-dimensional mixed time series feature map is obtained. Y H , the high-dimensional mixed time series feature map Y H Mixed time series feature plot Y After residual connection, a new high-dimensional mixed time series feature map is obtained Y HL , residual connections can effectively improve the generalization ability of the model;

[0059] New high-dimensional mixed time series feature map Y HL After two-dimensional batch normalization operation, input them into m The uniform sampling operation of different time scales is performed in the feature sequence mixers of different time scales, and then input into their respective MLP modules (referring to the MLP modules of the corresponding feature sequence mixers) for training, and finally output m Feature maps at different time scales Y HH .

[0060] High-dimensional mixed time series feature map Y HL Respectively m Feature maps at different time scales Y HHAfter the residual connection, the output of the mixer layer is obtained m Feature maps at different time scales Y M .

[0061] Preferably, the MLP module is a linear multilayer perceptron (MLP); the MLP module in the time series mixer consists of a fully connected layer, a ReLU activation function and a dropout layer for improving the generalization ability of the model, and the MLP module in the feature sequence mixer consists of a linear connection of five modules: a fully connected layer, a ReLU activation function, a dropout layer, a fully connected layer and a dropout layer.

[0062] Preferably, see Figure 4 The spatial feature attention layer first focuses on the mixed time series feature map Y' Perform sampling operations to generate mixed time series feature maps Y 1, , the mixed time series feature map Y 1 is input to the mean pooling layer on the first dimension, and then activated by the sigmoid function after training through the 1×1 convolution layer, the nonlinear pooling layer and the 1×1 convolution layer. Finally, the output layer is reconstructed to ensure that the shape and size of the output feature map are consistent with the mixed time series feature map of the input mixer layer. Y Consistent.

[0063] Furthermore, the spatial feature attention layer is used to calculate the mixed time series feature map Y' The sampling operation is as follows: taking the three-dimensional grid coordinate information (i.e. spatial position data) of a single wind turbine as the center, sampling the nearest K Mixed time series data of wind turbines to generate the sampled mixed time series feature graph Y 1, ,in K <N / 4, K is an integer.

[0064] The time domain projection layer outputs the feature map of the feature sequence mixer Y M After training, it is reconstructed into a predicted time series of the specified output sequence length; the core module of the time domain projection layer is a row-by-row fully connected layer, which not only learns the time features, but also can transform the original T The input time series of time series length is adjusted to the prediction length of the desired output prediction layer .

[0065] Specifically, Figure 5 As shown, this embodiment includes parallel arrangement mA time domain projection layer; each feature sequence mixer outputs a feature map Y M Input into the corresponding time domain projection layer respectively, each feature map Y M The transposition operation is first performed in the corresponding time domain projection layer, and then reconstructed into a predicted time series with the same specified output sequence length after training through the respective fully connected layers (referring to the fully connected layers of the corresponding time domain projection layer).

[0066] Output prediction layer, through the adder m The predicted time series of the specified output sequence length are added together to obtain the final predicted time series for output.

[0067] Preferably, m is a natural number greater than or equal to 2. In this embodiment, m=3, that is, this embodiment includes three time scales: normal time scale, short-term time scale and long-term time scale.

[0068] This embodiment also provides an offshore wind farm power prediction device, which includes a processor and a memory, in which the above-mentioned offshore wind farm power prediction method integrating spatiotemporal information is stored, and the processor is used to execute the above-mentioned offshore wind farm power prediction method integrating spatiotemporal information.

[0069] The prediction method of this embodiment integrates multiple time series data and wind turbine location information data, and is more reliable. It adopts a novel deep time series network TSMixer as the skeleton model, and improves the deep time series network TSMixer model by adding a spatial feature attention layer and feature sequence mixers at multiple time scales. In the offshore wind power prediction task, the coupling characteristics between wind turbines at multiple locations in the same wind farm and the feature information of the historical time series data of a single wind turbine at different time scales are considered, which improves the feature information extraction capability of the model and can more accurately predict the power of offshore wind farms.

[0070] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. For those skilled in the art, the present invention may have various modifications and variations. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. A method for predicting offshore wind farm power by integrating spatiotemporal information, characterized in that: include: Obtain mixed time series data of all wind turbines in the offshore wind farm to construct a tensor dataset, and divide the tensor dataset into a training set, a validation set, and a test set; The mixed time series data includes spatial position data of the wind turbine and multidimensional time series data; The multi-dimensional time series data includes time series data of wind speed, wind direction, air density, temperature, humidity, turbulence intensity and wind turbine power; wherein, For the The horizontal axis of the wind turbine is For the The vertical coordinate of a wind turbine is For the Height of blades of wind turbine; The TSMixer model is trained using the training set to obtain an offshore wind farm power prediction model; The validation set is used to adjust and verify the hyperparameters of the offshore wind farm power prediction model; Input the test set into the offshore wind farm power prediction model to realize offshore wind farm power prediction; The TSMixer model includes: The time series input layer is used to output the input mixed time series data as a mixed time series feature map Y ;in, , N is the total number of wind turbines in the offshore wind farm, T is the length of the mixed time series, C is the number of features of the mixed time series, represents Euclidean space; The mixer layer includes a time series mixer and a parallel setup m A feature sequence mixer of different time scales, wherein the MLP module in the time series mixer is cascaded with a spatial feature attention layer; the mixed time series feature map Y Input to the mixer layer and output is m Feature maps at different time scales Y M , m is a natural number greater than or equal to 2; The mixed time series feature map Y The processing flow in the mixer layer is: The time series mixer receives the mixed time series feature graph Y After that, a two-dimensional batch normalization operation and a transposition operation are performed to obtain a mixed time series feature map. Y T , mixed time series feature map Y T After feature extraction by the MLP module, a higher-dimensional mixed time series feature map is obtained. Y' , mixed time series feature map Y' After the spatial feature attention layer processing and transposition operation, a high-dimensional mixed time series feature map is obtained. Y H , the high-dimensional mixed time series feature map Y H Mixed time series feature plot Y After residual connection, a new high-dimensional mixed time series feature map is obtained Y HL ; New high-dimensional mixed time series feature map Y HL After two-dimensional batch normalization operation, input them into m The uniform sampling operation of different time scales is performed in the feature sequence mixer of different time scales, and then input into their respective MLP modules for training, and finally output m Feature maps at different time scales Y HH ; High-dimensional mixed time series feature map Y HL Respectively m Feature maps at different time scales Y HH After the residual connection, the output of the mixer layer is obtained m Feature maps at different time scales Y M ; The spatial feature attention layer first performs an attention on the mixed time series feature map. Y' Perform sampling operations to generate mixed time series feature maps Y 1, , the mixed time series feature map Y 1 is input to the mean pooling layer on the first dimension, and then activated by the sigmoid function after training through the 1×1 convolution layer, the nonlinear pooling layer and the 1×1 convolution layer. Finally, the output layer is reconstructed to ensure that the shape and size of the output feature map are consistent with the mixed time series feature map of the input mixer layer. Y consistent; among them K The number of wind turbines sampled, K <N / 4; The spatial feature attention layer is used to mix the time series feature map Y' The sampling operation is as follows: taking the location of a single wind turbine as the center, sampling the K Mixed time series data of wind turbines to generate the sampled mixed time series feature graph Y 1, ; Time domain projection layer, m Feature maps at different time scales Y M After training separately, reconstruct them into prediction time series with the same specified output sequence length; Output prediction layer, through the adder m The predicted time series of the specified output sequence length are added together to obtain the final predicted time series for output.

2. The offshore wind farm power prediction method integrating spatiotemporal information according to claim 1 is characterized in that: The spatial position data of the wind turbine is obtained by obtaining the distribution positions of all wind turbines in the offshore wind farm, plotting the positions of all wind turbines in the offshore wind farm on a two-dimensional coordinate system by grid division, performing two-dimensional position encoding on each grid, recording the horizontal and vertical coordinates and blade height of the wind turbine in the grid, and obtaining the spatial position data of the wind turbine.

3. The offshore wind farm power prediction method integrating spatiotemporal information according to claim 1 is characterized in that: Each feature sequence mixer outputs a feature map Y M Input into the corresponding time domain projection layer respectively, each feature map Y M The transposition operation is first performed in the corresponding time domain projection layer, and then reconstructed into a predicted time series with the same specified output sequence length after training through the respective fully connected layers.

4. The offshore wind farm power prediction method integrating spatiotemporal information according to claim 1 is characterized in that: The MLP module in the time series mixer consists of a fully connected layer, a ReLU activation function and a dropout layer; the MLP module in the feature sequence mixer consists of a linear connection of five modules: a fully connected layer, a ReLU activation function, a dropout layer, a fully connected layer and a dropout layer.

5. The offshore wind farm power prediction method integrating spatiotemporal information according to claim 1 is characterized in that: Randomly split the tensor dataset into training, validation, and test sets in proportion.

6. An offshore wind farm power prediction device integrating spatiotemporal information, characterized in that: The prediction device includes a processor and a memory, wherein the memory stores the offshore wind farm power prediction method integrating spatiotemporal information as described in any one of claims 1 to 5, and the processor is used to execute the offshore wind farm power prediction method integrating spatiotemporal information as described in any one of claims 1 to 5.

Images