A method and device of the prediction about photovoltaic generation and a method and device of the composition for the model used for the prediction about photovoltaic generation

Abstract

A method for establishing a solar power generation prediction model is implemented by a computing device. The computing device stores an initial solar power generation prediction model and multiple sets of sky image training data comprising multiple sky image groups. Based on the sky image groups, the computing device trains a contrastive learning feature module using a contrastive learning algorithm. Based on the sky image groups and multiple solar power generation markers, the computing device trains at least one convolutional module with fixed weights of the contrastive learning module. Furthermore, based on multiple contrastive learning feature vectors and multiple convolutional feature vectors output from the sky image groups via the contrastive learning module and the at least one convolutional module, and the solar power generation markers, a fully connected layer module is trained to obtain a trained solar power generation prediction model.

Claims

1. A method for establishing a solar power generation prediction model, implemented by a computing device, the computing device storing an initial solar power generation prediction model and multiple sets of sky image training data, the initial solar power generation prediction model comprising a contrastive learning feature module including an encoder, at least one convolutional module, and a fully connected layer module for connecting the contrastive learning feature module and the at least one convolutional module and including a fully connected layer, each convolutional module including a convolutional layer, a pooling layer and an activation function layer, each set of sky image training data comprising a set of sky images and a solar power generation label corresponding to the set of sky images, each set of sky images comprising at least one sky image, the method for establishing the solar power generation prediction model comprising the following steps: (A) Based on the sky image set, the encoder of the contrastive learning feature module is trained using a contrastive learning algorithm; and (B) Based on the sky image set and the solar power generation markers, with the weights of the contrastive learning module fixed, the convolutional layer and the activation function layer corresponding to each convolutional module are trained, and the fully connected layer of the fully connected layer module is trained based on the multiple contrastive learning feature vectors and multiple convolutional feature vectors output by the sky image set through the contrastive learning module and the at least one convolutional module, and the solar power generation markers, to obtain a trained solar power generation prediction model.

2. The method for establishing a solar power generation prediction model as described in claim 1, wherein, The contrastive learning feature module also includes a data augmentation function, and step (A) further includes the following sub-steps: (A-1) For each sky image group, the data augmentation function is used to obtain multiple augmented sky image groups corresponding to the sky image group; (A-2) Any of the sky image groups is taken as a representative sample group, and the augmented sky image groups corresponding to the representative sample group are taken as multiple positive sample groups of the representative sample group, and the sky image groups other than the representative sample group and their augmented sky image groups are taken as multiple negative sample groups of the representative sample group, and the representative sample group, the positive sample groups of the representative sample group and the negative sample groups of the representative sample group are taken as a single unlabeled image training data; and (A-3) The encoder is obtained using the contrastive learning algorithm based on all the unlabeled image training data.

3. The method for establishing a solar power generation prediction model as described in claim 2 further includes the following steps after step (B): (C) For each unlabeled image training data, a two-dimensional scatter map corresponding to the unlabeled image training data is obtained using a t-distribution random neighbor embedding method based on the unlabeled image training data and the contrastive learning algorithm.

4. The solar power generation prediction model establishment method as described in claim 1, wherein the computing device further stores multiple sets of sky images to be verified, each set of sky images to be verified contains at least one sky image to be verified, and after step (B), the method further includes the following steps: (D) For each sky image to be verified, a heat map corresponding to the set of sky images to be verified is obtained by using a gradient weighted activation mapping algorithm based on the sky image to be verified and the at least one convolution module.

5. The method for establishing a solar power generation prediction model as described in claim 1, wherein, Each sky image group also includes solar power generation features corresponding to at least one sky image. In step (A), the computing device trains the encoder of the contrast learning feature module using a contrast learning algorithm based on the sky images in the sky image group. In step (B), the computing device trains the convolutional layer and activation function layer corresponding to each convolutional module based on the sky images and solar power generation features in the sky image group, as well as the solar power generation labels, while fixing the weights of the contrast learning module. The computing device also trains the fully connected layer of the fully connected layer module based on the multiple contrast learning feature vectors output by the contrast learning module from the sky images in the sky image group, the multiple convolutional feature vectors output by the at least one convolutional module from the sky images and solar power generation features in the sky image group, and the solar power generation labels, to obtain a trained solar power generation prediction model.

6. A method for predicting solar power generation, implemented by a computing device storing a solar power generation prediction model for predicting the solar power generation of a set of sky images to be analyzed, the set of sky images to be analyzed including at least one sky image to be analyzed, the solar power generation prediction model including a contrastive learning feature module including an encoder, at least one convolutional module, and a fully connected layer module for connecting the contrastive learning feature module and the at least one convolutional module and including a fully connected layer, each convolutional module including a convolutional layer, a pooling layer and an activation function layer, the solar power generation prediction method comprising the following steps: (A) obtaining a contrastive learning feature vector corresponding to the set of sky images to be analyzed using the contrastive learning feature module based on a set of sky images to be analyzed including at least one sky image to be analyzed; (B) obtaining a convolutional feature vector corresponding to the set of sky images to be analyzed using the at least one convolutional module based on the set of sky images to be analyzed; and (C) obtaining a solar power generation using the fully connected layer module based on the contrastive learning feature vector and the convolutional feature vector.

7. The solar power generation prediction method as described in claim 6, wherein, In step (A), the group of sky images to be predicted also includes at least one pair of features of solar power generation to be predicted corresponding to at least one sky image to be predicted. The computing device obtains the contrast learning feature vector based on the sky images to be predicted in the group of sky images to be predicted using the contrast learning feature module.

8. A computing device for establishing a solar power generation prediction model, comprising a storage unit storing an initial solar power generation prediction model and multiple sets of sky image training data, the initial solar power generation prediction model comprising a contrastive learning feature module including an encoder, at least one convolutional module, a fully connected layer module for connecting the contrastive learning feature module and the at least one convolutional module and including a fully connected layer, each convolutional module including a convolutional layer, a pooling layer and an activation function layer, each set of sky image training data comprising a set of sky images and a solar power generation label corresponding to the set of sky images, each set of sky images comprising at least one sky image; and a processing unit electrically connected to the storage unit; wherein, The processing unit trains the encoder of the contrast learning feature module using a contrast learning algorithm based on the sky image set. Based on the sky image set and the solar power generation markers, the processing unit trains the convolutional layer and activation function layer corresponding to each convolutional module while fixing the weights of the contrast learning module. Based on the multiple contrast learning feature vectors and multiple convolutional feature vectors output by the sky image set through the contrast learning module and the at least one convolutional module, and the solar power generation markers, the fully connected layer of the fully connected layer module is trained to obtain a trained solar power generation prediction model.

9. The computing apparatus for establishing a solar power generation prediction model as described in claim 8, wherein, The contrastive learning feature module also includes a data augmentation function. For each sky image group, the processing unit uses the data augmentation function to obtain multiple augmented sky image groups corresponding to the sky image group. The processing unit takes any one of these sky image groups as a representative sample group, and takes the augmented sky image groups corresponding to the representative sample group as multiple positive sample groups of the representative sample group. The processing unit takes the sky image groups other than the representative sample group and their augmented sky image groups as multiple negative sample groups of the representative sample group. The processing unit takes the representative sample group, the positive sample groups of the representative sample group, and the negative sample groups of the representative sample group as a set of unlabeled image training data. The processing unit obtains the encoder based on all the unlabeled image training data using the contrastive learning algorithm.

10. The computing apparatus for establishing a solar power generation prediction model as described in claim 9, wherein after obtaining the trained solar power generation prediction model, for each unlabeled image training data, the processing unit obtains a two-dimensional scatter map corresponding to the unlabeled image training data using a t-distribution random neighbor embedding method based on the unlabeled image training data and the contrastive learning algorithm.

11. The computing apparatus for establishing a solar power generation prediction model as described in claim 8, wherein the storage unit further stores multiple sets of sky images to be verified, each set of sky images to be verified containing at least one sky image to be verified, and after obtaining the trained solar power generation prediction model, for each sky image to be verified, the processing unit obtains a heat map corresponding to the set of sky images to be verified using a gradient-weighted class activation mapping algorithm based on the sky image to be verified and the at least one convolutional module.

12. The computing apparatus for establishing a solar power generation prediction model as described in claim 8, wherein, Each sky image group also includes solar power generation features corresponding to at least one sky image. The processing unit trains the encoder of the contrast learning feature module using a contrast learning algorithm based on the sky images in the sky image group. The processing unit trains the convolutional layer and activation function layer corresponding to each convolutional module with fixed weights based on the sky images and solar power generation features in the sky image group, as well as the solar power generation labels. The processing unit trains the fully connected layer of the fully connected layer module based on the multiple contrast learning feature vectors output by the contrast learning module from the sky images in the sky image group, the multiple convolutional feature vectors output by the at least one convolutional module from the sky images and solar power generation features in the sky image group, as well as the solar power generation labels, to obtain a trained solar power generation prediction model.

13. A computing device for predicting solar power generation, comprising: a storage unit storing a solar power generation prediction model for predicting the solar power generation of one of a set of sky images to be analyzed, the set of sky images to be analyzed including at least one sky image to be analyzed, the solar power generation prediction model including a contrastive learning feature module including an encoder, at least one convolutional module, a fully connected layer module for connecting the contrastive learning feature module and the at least one convolutional module and including a fully connected layer, each convolutional module including a convolutional layer, a pooling layer and an activation function layer; and a processing unit electrically connected to the storage unit; wherein, The processing unit obtains a contrast learning feature vector corresponding to the sky image group to be predicted using the contrast learning feature module based on a sky image group to be predicted containing at least one sky image to be predicted. Based on the sky image group to be predicted, the processing unit obtains a convolutional feature vector corresponding to the sky image group to be predicted using the at least one convolution module. Based on the contrast learning feature vector and the convolutional feature vector, the processing unit obtains a solar power generation using the fully connected layer module.

14. The computing device for predicting solar power generation as described in claim 13, wherein, The group of sky images to be predicted also includes at least one pair of features of solar power generation to be predicted corresponding to at least one sky image to be predicted. The processing unit obtains the contrast learning feature vector based on the sky images to be predicted in the group of sky images to be predicted using the contrast learning feature module.

Images