A wind power prediction method based on deep learning fusion model

Abstract

The invention relates to the technical field of wind power prediction, and discloses a method for predicting wind power based on a deep learning fusion model, comprising: matching an operating power station whose similarity with environmental data and configuration data of a target power station is greater than a set similarity threshold; training a power model through the operating data of the operating power station; applying the branches of the trained power model to the operating data of the target power station for independent training, and after the branches of the power model converge, retraining the entire power model, adding a relational pattern reduction layer to the power model, the relational pattern reduction layer outputs a result representing the connection between objects, and the output loss of the relational pattern reduction layer is used as the loss of retraining; predicting the future short-term and medium-term wind power of each wind turbine generator of the target power station through the retrained power model; the method of the invention can shorten the window period of no predicted wind power to within half a month.

Description

Technical Field

[0001] The present invention relates to the technical field of wind power prediction, and more specifically, to a wind power prediction method based on a deep learning fusion model. Background Art

[0002] Wind power forecasts are also divided into ultra-short-term, short-term and medium-term. Wind power is mainly affected by wind speed, temperature and air pressure (temperature and air pressure affect air density). For power stations in operation, historical meteorological data such as wind speed and direction, temperature, air pressure and actual power can be provided.

[0003] Based on these historical data, the meteorological parameters of historical forecasts can be corrected using deep learning methods to obtain a correction model, which is used to correct the current 7-10 days of meteorological forecast parameters, and then the artificial intelligence model is trained based on the historical wind speed-temperature-pressure-power data to obtain the current short-term and medium-term forecast power. The first problem is that meteorological parameters are generally the overall expression of the meteorology of the entire power station area, but there will inevitably be differences in local meteorology, and the wind power of wind turbines is precisely affected by local meteorology. The current method is equivalent to a nonlinear fitting of the overall meteorology of a region to the overall wind power of the power station. Since the local meteorological differences are ignored, the generalization of the artificial intelligence model is poor. Another problem is that if local meteorological differences are considered, the difficulty of model convergence will be greatly increased. The artificial intelligence model relies on the meteorological data and actual power data generated by the operation of the power station for training, but for the power station that has been built not long ago, there is a lack of operation data, and the training of the artificial intelligence model cannot be carried out, resulting in a long window period without predicted wind power. Summary of the invention

[0004] The present invention provides a wind power prediction method based on a deep learning fusion model, which solves the technical problem in the related art that a power station that has been built not long ago lacks operating data and the training of an artificial intelligence model cannot be carried out, resulting in a long window period without predicted wind power.

[0005] The present invention provides a wind power prediction method based on a deep learning fusion model, comprising:

[0006] Collect environmental data and configuration data of the target power station;

[0007] Matching operating power plants whose similarity of environmental data and configuration data with the target power plant is greater than a set similarity threshold;

[0008] The power model is trained by operating data of the operating power station. The operating data of the operating power station needs to be structured to obtain structured data. The structured data includes structured data of two structural dimensions. The basic structural elements included in the first structural dimension are objects, and the basic structural elements included in the second structural dimension are time series units. The power model includes a trunk and branches. The trunk includes a first hidden layer, a second hidden layer, a mode fusion layer and a first output layer. The first hidden layer inputs the operating data of the operating power station and outputs a first hidden state to the second hidden layer. The second hidden layer outputs a second hidden state to the mode fusion layer. The mode fusion layer outputs a mode fusion state to the first output layer. The first output layer outputs the future short-term and medium-term wind power of each wind turbine generator of the target power station.

[0009] The branch includes a third hidden layer, a fourth hidden layer and a second output layer, wherein the third hidden layer inputs the operation data of the operation power station, outputs the third hidden state to the fourth hidden layer, the fourth hidden layer outputs the fourth hidden state to the mode fusion layer and the second output layer, and the second output layer outputs the future ultra-short-term wind power of each wind turbine generator of the target power station;

[0010] The branches of the trained power model are independently trained using the operating data of the target power plant. After the branches of the power model converge, the entire power model is trained again. A relational model restoration layer is added to the power model. The relational model restoration layer outputs the result representing the connection between objects. The output loss of the relational model restoration layer is used as the loss of retraining.

[0011] The future short-term and medium-term wind power of each wind turbine in the target power plant is predicted by the retrained power model.

[0012] Furthermore, the operating power station refers to a power station other than the target power station, and the continuous operation time is greater than 3 years.

[0013] Furthermore, when the power model is trained by operating data of the running power station, the losses of the outputs of the first output layer and the second output layer are back-propagated.

[0014] Furthermore, the method for structuring the operation data of the power station includes:

[0015] Extract objects that can be associated with specific features;

[0016] Extract specific features associated with the object, and divide the extracted specific feature sampling period into multiple parts, one part is mapped to a time sequence unit, and the order of the time sequence units is consistent with the time order of the sampling period of the specific feature;

[0017] Establishing a mapping between the extracted objects and their associated specific features;

[0018] For any two objects, if there is a correlation related to power prediction between the associated specific features or there is a correlation related to power prediction between the objects, an object relationship is established between the two objects.

[0019] Furthermore, the extracted objects that can be associated with specific features include: a power station, a wind turbine, a local area, and the entire area where the power station is located.

[0020] Further, the correlations related to power prediction include:

[0021] The local area is within the overall area where the power station is located;

[0022] The power station is located in the overall area where the power station is located;

[0023] The wind turbines are located in the overall area where the power station is located;

[0024] Wind turbines are located in local areas;

[0025] Wind turbines belong to power stations.

[0026] Furthermore, objects have specific characteristics associated with them as follows:

[0027] Specific characteristics associated with a power plant include: the actual power generated by the power plant;

[0028] The specific features associated with the wind turbine generator include: operating parameters of the wind turbine generator and the location coordinates of the wind turbine generator, and the operating parameters of the wind turbine generator include the actual power generated by the wind turbine generator;

[0029] The specific characteristics associated with the local area include: meteorological parameters of the local area;

[0030] The specific characteristics associated with the overall region where the power station is located include: meteorological parameters of the overall region.

[0031] Furthermore, the expression of the first hidden layer is as follows:

[0032]

[0033]

[0034]

[0035]

[0036] in, , , represents the first, second, and third weight parameters, , , represents the first, second and third bias parameters, represents the dot product, , and Respectively represent the tth first, second, and third intermediate states of the vth object, where The part of the specific feature associated with the vth object corresponding to the tth time unit, and They represent the t-th and t-1-th first hidden states of the v-th object, respectively. n ≥ t ≥ 1, where n represents the total number of data items of the one-dimensional structured data. When t = 1 , tanh is the hyperbolic tangent function, Represents the sigmoid function;

[0037] The expression of the second hidden layer is as follows:

[0038]

[0039]

[0040]

[0041] in represents the second hidden state of the vth object, , , and Respectively represent the nth first hidden state of the vth object, represents the attention coefficient of the vth and uth units, represents the attention coefficient of the vth and sth units, represents the normalized attention coefficients of the vth and uth units, represents a set of objects that have an association with the vth object related to power prediction, represents an exponential function with a natural constant as base, represents the trainable attention vector parameter, CONCAT represents the concatenation operation, represents transpose, represents the first recognition weight parameter, LeakyReLU is the LeakyReLU activation function;

[0042] The expression of the third hidden layer is as follows:

[0043]

[0044] in and denote the third hidden states of the t-th and t-1-th sequential units of the v-th object, respectively. The part corresponding to the t-th temporal unit that represents the specific feature associated with the v-th object, and are the fourth and fifth weight parameters, is the fourth bias parameter, tanh is the hyperbolic tangent function;

[0045] The expression of the fourth hidden layer is as follows:

[0046]

[0047] in represents the fourth hidden state of the vth object, is the set of cells of the data matrix that are associated with the vth cell of the data matrix, is a learnable scalar parameter, represents the second recognition weight parameter, MLP represents a multi-layer perceptron;

[0048] The expression of the mode fusion layer is as follows:

[0049]

[0050] in represents the mode fusion state matrix, a row vector of which represents the mode fusion state of an object. represents the second state matrix, a row vector of which represents the second hidden state of an object, represents the fourth state matrix, a row vector of which represents the fourth hidden state of an object, represents multidimensional noise, The identity matrix representing the multidimensional standard Gaussian distribution;

[0051] The expression of the relational model reduction layer is as follows:

[0052]

[0053] in represents the relational pattern matrix, Indicates that a value greater than 0.5 is changed to 1. The element in the i-th row and j-th column of the relationship pattern matrix indicates whether there is a correlation related to power prediction between the i-th object and the j-th object. If the element value is 1, it means there is a correlation, otherwise it does not exist. Indicates transpose.

[0054] Furthermore, the expression of the first output layer is as follows:

[0055]

[0056] in represents a first output vector, one component of which represents the short-term or medium-term wind power of a wind turbine, represents the second output weight parameter, represents the second output bias parameter;

[0057] The expression of the second output layer is as follows:

[0058]

[0059] in represents the second output vector, one component of which represents the ultra-short-term wind power of a wind turbine, represents the second output weight parameter, Represents the second output bias parameter.

[0060] The present invention provides a computer-readable medium for storing computer-readable instructions, which, when read by a computer, can execute the steps of the aforementioned wind power prediction method based on a deep learning fusion model.

[0061] The beneficial effects of the present invention are:

[0062] Different from general transfer learning, the present invention adds a branch of ultra-short-term power prediction in the power model to supervise the training of the trunk used to predict medium-term or ultra-long-term wind power prediction. The branch of ultra-short-term power prediction can use the data of the initial operation of the power plant for training convergence, and then assist the trunk to adjust parameters to match the target power plant through the training method of the results of the connection between objects, thereby shortening the window period of no predicted wind power to within half a month. BRIEF DESCRIPTION OF THE DRAWINGS

[0063] Figure 1 is a flow chart of a wind power prediction method based on a deep learning fusion model of the present invention;

[0064] Figure 2 This is a schematic diagram of the power model of the present invention. Figure 1 ;

[0065] Figure 3 This is a schematic diagram of the power model of the present invention. Figure 2 . DETAILED DESCRIPTION

[0066] The subject matter described herein will now be discussed with reference to example implementations. It should be understood that the discussion of these implementations is only to enable those skilled in the art to better understand and implement the subject matter described herein, and the functions and arrangements of the elements discussed may be changed without departing from the scope of protection of the present specification. Various examples may omit, replace, or add various processes or components as needed. In addition, the features described in some examples may also be combined in other examples.

[0067] At least one embodiment of the present invention discloses a wind power prediction method based on a deep learning fusion model, such as Figure 1 As shown, the following steps are included:

[0068] Step 100, collecting environmental data and configuration data of the target power station;

[0069] Step 200, matching the operating power station whose similarity of environmental data and configuration data with the target power station is greater than a set similarity threshold;

[0070] Specifically, the similarity of the aforementioned content can be measured by using general similarity metrics such as cosine distance and Euclidean distance. Of course, when measuring by these methods, the environmental data and configuration data of the target power station can be generated through feature engineering to represent the features of the vector, and the environmental data and configuration data of the operating power station can be generated through feature engineering to represent the features of the vector using the same method.

[0071] Specifically, in the above content, the operating power station refers to a power station other than the target power station, and the continuous operation time is greater than 3 years.

[0072] Specifically, the environmental data includes the latitude and longitude of the target power station and the geographical space data within 5 kilometers around it;

[0073] Specifically, the configuration data includes the number of wind turbines of the target power station, the power generation scale, the density of wind turbines, and the area occupied by the power station.

[0074] Step 300, the power model is trained by running the power station's operating data (such as wind speed and direction, temperature, air pressure and other meteorological data and actual power). The power station's operating data needs to be structured to obtain structured data. The structured data includes structured data of two structural dimensions. The basic structural elements included in the first structural dimension are objects, and the basic structural elements included in the second structural dimension are time series units; Figure 2As shown, the power model includes a trunk and branches, the trunk includes a first hidden layer, a second hidden layer, a mode fusion layer and a first output layer, wherein the first hidden layer inputs the operation data of the operating power station, outputs the first hidden state to the second hidden layer, the second hidden layer outputs the second hidden state to the mode fusion layer, the mode fusion layer outputs the mode fusion state to the first output layer, and the first output layer outputs the future short-term and medium-term wind power of each wind turbine generator of the target power station;

[0075] The branch includes a third hidden layer, a fourth hidden layer and a second output layer, wherein the third hidden layer inputs the operation data of the operation power station, outputs the third hidden state to the fourth hidden layer, the fourth hidden layer outputs the fourth hidden state to the mode fusion layer and the second output layer, and the second output layer outputs the future ultra-short-term wind power of each wind turbine generator of the target power station;

[0076] Step 400, the branches of the power model trained in Step 300 are independently trained using the operating data of the target power plant. After the branches of the power model converge, the entire power model is trained again;

[0077] like Figure 3 As shown, during the process of retraining the entire power model, a relational model restoration layer is added to the power model, and the relational model restoration layer outputs results representing the connections between objects.

[0078] Because the medium-term cycle is generally hundreds of times longer than the ultra-short term, hundreds of ultra-short term training samples will be obtained in the process of obtaining a medium-term training sample. When the number of medium-term training samples is small, the ultra-short term training samples are sufficient to support the training of the power model.

[0079] Unsupervised training is then performed through the connections between known objects to update the parameters of the power model. Supervised and unsupervised training is combined to train a suitable power model based on the operating data of the target power station within a short period of time (three days to one week) when it is initially started up to predict the future short-term and medium-term wind power of each wind turbine in the target power station, providing a basis for the precise control of the power generation of the power station to each wind turbine.

[0080] Furthermore, the method does not rely on a large amount of historical data of short-term and medium-term wind power during the operation of the target power station for training, and can play a timely role in the early stage of the operation of the target power station.

[0081] Specifically, the method for structuring the operation data of the power station includes:

[0082] Extract objects that can be associated with specific features;

[0083] Extract specific features associated with the object, and divide the extracted specific feature sampling period into multiple parts, one part is mapped to a time sequence unit, and the order of the time sequence units is consistent with the time order of the sampling period of the specific feature;

[0084] Establishing a mapping between the extracted objects and their associated specific features;

[0085] For any two objects, if there is a correlation related to power prediction between the associated specific features or there is a correlation related to power prediction between the objects, an object relationship is established between the two objects;

[0086] In some embodiments of the present invention, the extracted objects capable of being associated with specific features include: a power station, a wind turbine, a local area, and the entire area where the power station is located;

[0087] In some embodiments of the invention, objects have specific features associated with them as follows:

[0088] Specific characteristics associated with a power plant include: the actual power generated by the power plant;

[0089] The specific features associated with the wind turbine generator include: operating parameters of the wind turbine generator and the location coordinates of the wind turbine generator, and the operating parameters of the wind turbine generator include the actual power generated by the wind turbine generator;

[0090] The specific characteristics associated with the local area include: meteorological parameters of the local area;

[0091] The specific characteristics associated with the overall region where the power station is located include: meteorological parameters of the overall region.

[0092] In some embodiments of the present invention, the correlations related to power prediction include:

[0093] The local area is within the overall area where the power station is located;

[0094] The power station is located in the overall area where the power station is located;

[0095] The wind turbines are located in the overall area where the power station is located;

[0096] Wind turbines are located in local areas;

[0097] Wind turbines belong to power stations.

[0098] Specifically, the above content selects operating data of one or more operating power stations to train the power model.

[0099] The above content is specific, ultra-short-term forecast: wind power forecast for the next 15 minutes to 1 hour.

[0100] Short-term forecast: Wind power forecast for the next 1 to 7 days.

[0101] Medium to long-term forecast: Wind power forecast for the next 8 to 30 days.

[0102] Step 500, predicting the future short-term and medium-term wind power of each wind turbine of the target power station through the retrained power model.

[0103] Specifically, the expression of the first hidden layer is as follows:

[0104]

[0105]

[0106]

[0107]

[0108] in, , , represents the first, second, and third weight parameters, , , represents the first, second and third bias parameters, represents the dot product, , and Respectively represent the tth first, second, and third intermediate states of the vth object, where The part of the specific feature associated with the vth object corresponding to the tth time unit, and They represent the t-th and t-1-th first hidden states of the v-th object, respectively. n ≥ t ≥ 1, where n represents the total number of data items of the one-dimensional structured data. When t = 1 , tanh is the hyperbolic tangent function, Represents the sigmoid function.

[0109] The expression of the second hidden layer is as follows:

[0110]

[0111]

[0112]

[0113] in represents the second hidden state of the vth object, , , and Respectively represent the nth first hidden state of the vth object, represents the attention coefficient of the vth and uth units, represents the attention coefficient of the vth and sth units, represents the normalized attention coefficients of the vth and uth units, represents a set of objects that have an association with the vth object related to power prediction, represents an exponential function with a natural constant as base, represents the trainable attention vector parameter, CONCAT represents the concatenation operation, represents transpose, represents the first identification weight parameter, and LeakyReLU is the LeakyReLU activation function.

[0114] Specifically, the expression of the third hidden layer is as follows:

[0115]

[0116] in and denote the third hidden states of the t-th and t-1-th sequential units of the v-th object, respectively. The part corresponding to the t-th temporal unit that represents the specific feature associated with the v-th object, and are the fourth and fifth weight parameters, is the fourth bias parameter, and tanh is the hyperbolic tangent function.

[0117] The expression of the fourth hidden layer is as follows:

[0118]

[0119] in represents the fourth hidden state of the vth object, is the set of cells of the data matrix that are associated with the vth cell of the data matrix, is a learnable scalar parameter, represents the second recognition weight parameter, and MLP represents a multi-layer perceptron.

[0120] Specifically, the expression of the mode fusion layer is as follows:

[0121]

[0122] in represents the mode fusion state matrix, a row vector of which represents the mode fusion state of an object. represents the second state matrix, a row vector of which represents the second hidden state of an object, represents the fourth state matrix, a row vector of which represents the fourth hidden state of an object, represents multidimensional noise, The identity matrix representing the multidimensional standard Gaussian distribution.

[0123] Specifically, the expression of the first output layer is as follows:

[0124]

[0125] in represents a first output vector, one component of which represents the short-term or medium-term wind power of a wind turbine, represents the second output weight parameter, Represents the second output bias parameter.

[0126] The expression of the second output layer is as follows:

[0127]

[0128] in represents the second output vector, one component of which represents the ultra-short-term wind power of a wind turbine, represents the second output weight parameter, Represents the second output bias parameter.

[0129] Specifically, the expression of the relational model reduction layer is as follows:

[0130]

[0131] in represents the relational pattern matrix, Indicates that a value greater than 0.5 is changed to 1. The element in the i-th row and j-th column of the relationship pattern matrix indicates whether there is a correlation related to power prediction between the i-th object and the j-th object. If the element value is 1, it means there is a correlation, otherwise it does not exist. Indicates transpose.

[0132] At least one embodiment of the present invention discloses a computer-readable medium for storing computer-readable instructions, which, when read by a computer, can execute the steps of the aforementioned wind power prediction method based on a deep learning fusion model.

[0133] The above describes an embodiment of the present invention, but this embodiment is not limited to the above-mentioned specific implementation mode. The above-mentioned specific implementation mode is merely illustrative and not restrictive. Under the guidance of this embodiment, ordinary technicians in this field can also make more forms of equivalent embodiments, all of which are within the protection of this embodiment.

Claims

1. A wind power prediction method based on deep learning fusion model, characterized in that: include: Collect environmental data and configuration data of the target power station; Matching operating power plants whose similarity of environmental data and configuration data with the target power plant is greater than a set similarity threshold; The power model is trained by operating data of the operating power station. The operating data of the operating power station needs to be structured to obtain structured data. The structured data includes structured data of two structural dimensions. The basic structural elements included in the first structural dimension are objects, and the basic structural elements included in the second structural dimension are time series units. The power model includes a trunk and branches. The trunk includes a first hidden layer, a second hidden layer, a mode fusion layer and a first output layer. The first hidden layer inputs the operating data of the operating power station and outputs a first hidden state to the second hidden layer. The second hidden layer outputs a second hidden state to the mode fusion layer. The mode fusion layer outputs a mode fusion state to the first output layer. The first output layer outputs the future short-term and medium-term wind power of each wind turbine generator of the target power station. The branch includes a third hidden layer, a fourth hidden layer and a second output layer, wherein the third hidden layer inputs the operation data of the operation power station, outputs the third hidden state to the fourth hidden layer, the fourth hidden layer outputs the fourth hidden state to the mode fusion layer and the second output layer, and the second output layer outputs the future ultra-short-term wind power of each wind turbine generator of the target power station; The branches of the trained power model are independently trained using the operating data of the target power plant. After the branches of the power model converge, the entire power model is trained again. A relational pattern reduction layer is added to the power model. The relational pattern reduction layer outputs a relational pattern matrix representing the results of the connection between objects. The element in the i-th row and j-th column of the relational pattern matrix represents whether there is a correlation related to power prediction between the i-th object and the j-th object. If the element value is 1, it means there is a correlation, otherwise it means there is no correlation. The output loss of the relational pattern reduction layer is used as the loss of retraining. The future short-term and medium-term wind power of each wind turbine in the target power plant is predicted by the retrained power model.

2. The method for predicting wind power based on a deep learning fusion model according to claim 1, characterized in that: An operating power station refers to a power station other than the target power station, and has been in continuous operation for more than 3 years.

3. The method for predicting wind power based on a deep learning fusion model according to claim 1, characterized in that: When training the power model by running the operating data of the power plant, the losses of the outputs of the first output layer and the second output layer are back-propagated.

4. The method for predicting wind power based on a deep learning fusion model according to claim 1, characterized in that: Methods for structuring the operating data of an operating power plant include: Extract objects that can be associated with specific features; Extract specific features associated with the object, and divide the extracted specific feature sampling period into multiple parts, one part is mapped to a time sequence unit, and the order of the time sequence units is consistent with the time order of the sampling period of the specific feature; Establishing a mapping between the extracted objects and their associated specific features; For any two objects, if there is a correlation related to power prediction between the associated specific features or there is a correlation related to power prediction between the objects, an object relationship is established between the two objects.

5. A wind power prediction method based on deep learning fusion model according to claim 4, characterized in that: The extracted objects that can be associated with specific features include: power stations, wind turbines, local areas, and the overall area where the power stations are located.

6. A wind power prediction method based on deep learning fusion model according to claim 5, characterized in that: Relevance to power prediction includes: The local area is within the overall area where the power station is located; The power station is located in the overall area where the power station is located; The wind turbines are located in the overall area where the power station is located; Wind turbines are located in local areas; Wind turbines belong to power stations.

7. The method for predicting wind power based on a deep learning fusion model according to claim 5, characterized in that: Objects have specific characteristics associated with them: Specific characteristics associated with a power station include: the actual power generated by the power station; The specific features associated with the wind turbine generator include: operating parameters of the wind turbine generator and the location coordinates of the wind turbine generator, and the operating parameters of the wind turbine generator include the actual power generated by the wind turbine generator; The specific characteristics associated with the local area include: meteorological parameters of the local area; The specific characteristics associated with the overall region where the power station is located include: meteorological parameters of the overall region.

8. The method for predicting wind power based on a deep learning fusion model according to claim 1, characterized in that: The first hidden layer is expressed as follows: ； ； ； ； in, , , represents the first, second, and third weight parameters, , , represents the first, second and third bias parameters, represents the dot product, , and Respectively represent the tth first, second, and third intermediate states of the vth object, where The part of the specific feature associated with the vth object corresponding to the tth time unit, and They represent the t-th and t-1-th first hidden states of the v-th object, respectively. n ≥ t ≥ 1, where n represents the total number of data items of the one-dimensional structured data. When t = 1 , tanh is the hyperbolic tangent function, Represents the sigmoid function; The expression of the second hidden layer is as follows: ； ； ； in represents the second hidden state of the vth object, , , and denote the nth first hidden state of the vth object and the uth object respectively, represents the attention coefficient of the vth and uth objects, represents the attention coefficient of the vth and sth objects, represents the normalized attention coefficient of the vth and uth objects, represents a set of objects that have an association with the vth object related to power prediction, represents an exponential function with a natural constant as base, represents the trainable attention vector parameter, CONCAT represents the concatenation operation, represents transpose, represents the first recognition weight parameter, LeakyReLU is the LeakyReLU activation function; The expression of the third hidden layer is as follows: ； in and denote the third hidden states of the t-th and t-1-th sequential units of the v-th object, respectively. The part corresponding to the t-th temporal unit that represents the specific feature associated with the v-th object, and are the fourth and fifth weight parameters, is the fourth bias parameter, tanh is the hyperbolic tangent function; The expression of the fourth hidden layer is as follows: ； in represents the fourth hidden state of the vth object, is the set of data matrix objects that are associated with the vth object of the data matrix, is a learnable scalar parameter, represents the second recognition weight parameter, MLP represents a multi-layer perceptron; The expression of the mode fusion layer is as follows: ； in represents the mode fusion state matrix, a row vector of which represents the mode fusion state of an object. represents the second state matrix, a row vector of which represents the second hidden state of an object, represents the fourth state matrix, a row vector of which represents the fourth hidden state of an object, represents multidimensional noise, The identity matrix representing the multidimensional standard Gaussian distribution; The expression of the relational model reduction layer is as follows: ； in represents the relational pattern matrix, It means changing the value greater than 0.5 to 1. Indicates transpose.

9. A wind power prediction method based on deep learning fusion model according to claim 8, characterized in that: The expression of the first output layer is as follows: ； in represents a first output vector, one component of which represents the short-term or medium-term wind power of a wind turbine, represents the second output weight parameter, represents the second output bias parameter; The expression of the second output layer is as follows: ； in represents the second output vector, one component of which represents the ultra-short-term wind power of a wind turbine, represents the second output weight parameter, Represents the second output bias parameter.

10. A computer-readable medium, characterized in that It is used to store computer-readable instructions, which, when read by a computer, can execute the steps of a wind power prediction method based on a deep learning fusion model as described in any one of claims 1-9.

Images