Photovoltaic power prediction method and system based on multi-stage time sequence feature mining

Abstract

The invention discloses a photovoltaic power prediction method and system based on multi-stage time sequence feature mining, and the method comprises the steps: carrying out the feature extraction of a historical photovoltaic power sequence through an LSTM model, and obtaining a power time sequence feature; performing feature extraction on the historical satellite cloud picture sequence through a 3DCNN model to obtain cloud layer space-time dynamic features; performing feature fusion on the power time sequence feature and the cloud layer space-time dynamic feature based on a residual cross attention mechanism to obtain a multi-modal fusion feature; further mining time sequence characteristics in the fusion characteristics by using a Transform model to obtain deep multi-modal time sequence characteristics; and outputting a prediction result through the multi-layer perceptron to predict the photovoltaic power in the next 1-4 hours. The photovoltaic power prediction precision can be improved, support is provided for safe and stable operation of a power grid, and the method has certain engineering practical value.

Description

Technical Field

[0001] The present invention belongs to the technical field of photovoltaic power generation, and in particular relates to a photovoltaic power prediction method and system based on multi-stage time series feature mining. Background Art

[0002] Solar energy has attracted global attention due to its wide distribution, easy mining, efficient energy utilization and pollution-free characteristics, and has become a focus of renewable energy research. Photovoltaic power generation has become the best solar energy application method due to its advantages such as quietness, cleanliness, high geographical flexibility and short construction period. However, photovoltaic power generation is affected by meteorological factors such as sunshine, cloud cover and temperature, and its output power is intermittent and volatile, which poses a challenge to the stability of the power grid and restricts its widespread deployment. Therefore, it is crucial to accurately predict the power of photovoltaic power generation.

[0003] In the field of photovoltaic power prediction, the application of satellite cloud images plays a vital role. It provides real-time, large-scale weather conditions, including cloud coverage, distribution and movement, which is crucial for predicting short-term changes in photovoltaic power. There are very obvious time series characteristics between photovoltaic power series and satellite cloud image series. However, the existing prediction methods use insufficient time series feature extraction and ignore the coupling and redundancy between power features and image features, which limits the improvement of photovoltaic power prediction performance. How to make full use of the time series characteristics of historical data, explore the coupling characteristics between power features and image features, and further improve the performance of photovoltaic power prediction models is the current difficulty in improving the utilization of solar energy. Summary of the invention

[0004] The purpose of the present invention is to provide a photovoltaic power prediction method and system based on multi-stage time series feature mining to solve the problems raised in the above background technology.

[0005] To achieve the above object, the present invention provides the following technical solution: a photovoltaic power prediction method based on multi-stage time series feature mining, comprising the following steps:

[0006] Step 1: Extract features of historical photovoltaic power sequences through the LSTM model to obtain power time series features. Extract features of historical photovoltaic power sequences through the LSTM model to obtain power time series features. The process can be expressed as:

[0007]

[0008] Among them, F power is the power timing characteristic, f LSTM is the LSTM model, m is the number of LSTM layers used, [p 1 ,p 2 ,...,p n] is the historical photovoltaic power series, n is the length of the time series;

[0009] Step 2: Use the 3DCNN model to extract features from the historical satellite cloud image sequence to obtain the spatiotemporal dynamic characteristics of the cloud layer. The process can be expressed as: image =f 3DCNN ([i 1 ,i 2 ,...,i n ]);

[0010] Among them, F image is the power timing characteristic, f 3DCNN For the 3DCNN model, [i 1 ,i 2 ,...,i n ] is a historical satellite cloud image sequence; 3DCNN can be expressed by the following formula:

[0011] h (l) =f (l) (w (l) *h (l-1) +b (l) );

[0012] Step 3: Based on the residual cross-attention mechanism, the power time series features and the cloud spatiotemporal dynamic features are fused to obtain multimodal fusion features. The specific process is as follows:

[0013]

[0014] Step 4: Use the Transformer model to further explore the temporal characteristics in the fusion features and obtain deep multimodal temporal features. The specific process is as follows:

[0015]

[0016] Step 5: The prediction results are output through the multi-layer perceptron to predict the photovoltaic power in the next 1 to 4 hours. Based on the power time series characteristics, cloud image spatiotemporal characteristics and deep time series characteristics, as well as the descriptions of steps 2 and 3 above, the power characteristics, image characteristics and fusion characteristics can be obtained from the historical data. Finally, by re-learning these three characteristics, the prediction result y of the photovoltaic power can be obtained:

[0017] y=f MLP ([F power ,F image ,F final ]);

[0018] Among them, F MLP It is a multi-layer perceptron model.

[0019] Preferably, in step 1, LSTM can be expressed by the following formula:

[0020] i t =σ(W xi *x t +W hi *h t-1 +b i )

[0021] f t =σ(W xf *x t +W hf *h t-1 +b f )

[0022]

[0023] g t =σ(W xg *x t +W hg *h t-1 +b g )

[0024]

[0025] Among them, x t ,h t , c t They represent the input value, output value, and unit state of LSTM at time t respectively; h t-1 is the output value of LSTM at time t-1, i t , f t , g t They are input gate, forget gate and output gate, which are used to control the flow of information; W xi , W hi , b i , W xf , W hf , b f , W xc , W hc , b c , W xg , W hg , b g All represent learnable parameters, which enable the network to effectively process sequence data and have the ability to remember long-term dependencies. Tanh is the hyperbolic tangent function, σ(·) is the activation function, and “*” is the ordinary product of matrices. represents the Hadamard product.

[0026] Preferably, in step 2, * represents a convolution operation, w (l) is the weight of the lth convolutional layer, b(l) is the bias of the lth convolutional layer, f (l) is the activation function of the lth convolutional layer, h (l) ,h (l-1) are the outputs of the lth and l-1th convolutional layers respectively; in addition, a three-dimensional maximum pooling operation is used to downsample the feature map to reduce the amount of model calculation and the number of parameters while preventing overfitting. The process can be expressed as:

[0027]

[0028] Among them, Y is the output feature map, (x, y, z) is the position index in the output feature map, g is the input feature map of the pooling operation, M, N, O represent the size of the pooling window in height, depth and width respectively, S d , S h , S w are the moving steps of the pooling window in depth, height and width, respectively, and m, n, o are the index offsets corresponding to the current position of the pooling window.

[0029] Preferably, in step three, F fusion is a multimodal fusion feature, RS-Attention is a residual cross attention mechanism, Q, K, and V are query vectors, key vectors, and value vectors, respectively. Q is obtained through nonlinear mapping of power features, and K and V are obtained through nonlinear mapping of image features. k is the dimension of the key vector, and softmax is used to calculate the cross attention mechanism weights.

[0030] Preferably, in step 4, F final is the multimodal time series feature, f Transformer is the Transformer model, is the multimodal fusion feature at each historical moment, and n is the dimension of the fusion feature.

[0031] To achieve the above object, the present invention also provides the following technical solutions:

[0032] A photovoltaic power prediction system based on multi-stage time series feature mining includes a power feature extraction module, an image feature extraction module, a feature fusion module, a deep feature extraction module and a time series prediction module, which are used to execute the above-mentioned photovoltaic power prediction method based on multi-stage time series feature mining.

[0033] The power feature extraction module extracts features from the historical photovoltaic power sequence through the LSTM model to obtain power time series features;

[0034] The image feature extraction module extracts features from historical satellite cloud image sequences through a 3DCNN model to obtain the spatiotemporal dynamic characteristics of the cloud layer;

[0035] The feature fusion module fuses the power time series features and the cloud spatiotemporal dynamic features based on the residual cross-attention mechanism to obtain multimodal fusion features;

[0036] The deep feature extraction module uses the Transformer model to further explore the temporal characteristics in the fusion features and obtain deep multimodal temporal features;

[0037] The time series prediction module takes the significant spatiotemporal features of the image and the historical photovoltaic power as input, and outputs the photovoltaic power results for the next 1 to 4 hours through a multi-layer perceptron.

[0038] To achieve the above object, the present invention also provides the following technical solutions:

[0039] A computer device comprising:

[0040] one or more processors;

[0041] Memory;

[0042] And one or more programs, wherein the one or more programs are stored in the memory and are configured to be executed by the one or more processors, and when the programs are executed by the processors, the steps of the photovoltaic power prediction method based on multi-stage time series feature mining as described above are implemented.

[0043] Preferably, when the computer program is executed by a processor, the steps of the photovoltaic power prediction method based on multi-stage time series feature mining as described above are implemented.

[0044] Compared with the prior art, the present invention has the following beneficial effects:

[0045] (1) LSTM and 3DCNN are used to extract power time series features and image spatiotemporal features respectively. The deep time series characteristics of the fused features are further mined based on the Transformer model. The time series characteristics of historical data are captured in multiple stages, and the model's ability to represent the time series characteristics of multi-source data is successfully improved.

[0046] (2) The residual cross-attention mechanism is used to fuse the power features and image features, explore the coupling characteristics between the two features, reduce the redundancy of the fused features, and improve the availability of feature information. BRIEF DESCRIPTION OF THE DRAWINGS

[0047] Figure 1 It is a schematic diagram of the photovoltaic power prediction method of the present invention;

[0048] Figure 2A schematic diagram of a long short-term memory neural network used in the present invention;

[0049] Figure 3 A schematic diagram of the Transformer model used in the present invention;

[0050] Figure 4 Schematic diagram of the residual cross attention mechanism proposed in the present invention;

[0051] Figure 5 It is a schematic diagram of fitting the prediction results using the proposed method in an embodiment of the present invention. DETAILED DESCRIPTION

[0052] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only 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.

[0053] The present invention provides Figure 1 A technical solution shown: a photovoltaic power prediction method based on multi-stage time series feature mining, comprising the following steps:

[0054] Step 1: Extract features of historical photovoltaic power sequences through the LSTM model to obtain power time series features;

[0055] Step 2: Use the 3DCNN model to extract features from the historical satellite cloud image sequence to obtain the spatiotemporal dynamic characteristics of the cloud layer;

[0056] Step 3: Based on the residual cross-attention mechanism, the power time series features and the cloud spatiotemporal dynamic features are fused to obtain multimodal fusion features;

[0057] Step 4: Use the Transformer model to further explore the temporal characteristics in the fusion features and obtain deep multimodal temporal features;

[0058] Step 5: Output the prediction results through the multi-layer perceptron to predict the photovoltaic power in the next 1 to 4 hours.

[0059] Specifically, in the embodiment of the present invention, in step 1, power timing feature extraction is performed by the following method;

[0060] The LSTM model is used to extract the features of the historical photovoltaic power sequence to obtain the power time series features. The process can be expressed as:

[0061]

[0062] Among them, F power is the power time series feature, m is the number of LSTM layers used, [p 1 ,p 2 ,...,p n ] is the historical photovoltaic power series, n is the length of the time series; Figure 2 As shown, LSTM can be expressed by the following formula:

[0063] i t =σ(W xi *x t +W hi *h t-1 +b i )

[0064] f t =σ(W xf *x t +W hf *h t-1 +b f )

[0065]

[0066] g t =σ(W xg *x t +W hg *h t-1 +b g )

[0067]

[0068] Among them, x t ,h t , c t They represent the input value, output value, and unit state of LSTM at time t respectively; h t-1 is the output value of LSTM at time t-1, i t , f t , g t They are input gate, forget gate and output gate, which are used to control the flow of information; W xi , W hi , b i , W xf , W hf , b f , W xc , W hc , b c , W xg , W hg , b g Both represent learnable parameters, which enable the network to effectively process sequence data and have the ability to remember long-term dependencies; σ(·) is the activation function, represents the Hadamard product.

[0069] In step 2, the present invention uses a 3DCNN model to extract features from the historical satellite cloud image sequence to obtain the spatiotemporal dynamic features of the cloud layer. The process can be expressed as follows:

[0070] F image =f 3DCNN ([i 1 ,i 2 ,...,i n ])

[0071] Among them, F image is the power timing characteristic, f 3DCNN For the 3DCNN model, [i 1 ,i 2 ,...,i n ] is a historical satellite cloud image sequence; 3DCNN can be expressed by the following formula:

[0072] h (l) =f (l) (w (l) *h (l-1) +b (l) );

[0073] Among them, * represents the convolution operation, w (l) is the weight of the lth convolutional layer, b (l) is the bias of the lth convolutional layer, f (l) is the activation function of the lth convolutional layer, h (l) ,h (l-1) are the outputs of the lth and l-1th convolutional layers respectively; in addition, a three-dimensional maximum pooling operation is used to downsample the feature map to reduce the amount of model calculation and the number of parameters while preventing overfitting. The process can be expressed as:

[0074]

[0075] Among them, Y is the output feature map, (x, y, z) is the position index in the output feature map, g is the input feature map of the pooling operation, M, N, O represent the size of the pooling window in height, depth and width respectively, S d , S h , S w are the moving steps of the pooling window in depth, height and width, respectively; m, n, o are the index offsets corresponding to the current position of the pooling window; at the same time, the Batch Normalization operation is applied to reprocess the output feature map to speed up the convergence of the model and improve the stability of the model.

[0076] like Figure 3As shown in the figure, in step 3, the power time series features and cloud spatiotemporal dynamic features are fused through the residual cross attention mechanism to obtain multimodal fusion features. First, the key feature information in the cloud spatiotemporal features is screened by calculating the correlation between the power time series features and the cloud spatiotemporal features; secondly, the key cloud spatiotemporal features and the power time series features are merged through the splicing operation to obtain the fusion features. The specific process of the residual cross attention mechanism is:

[0077]

[0078] Among them, F fusion is a multimodal fusion feature, RS-Attention is a residual cross attention mechanism, Q, K, and V are query vectors, key vectors, and value vectors, respectively. Q is obtained through nonlinear mapping of power features, and K and V are obtained through nonlinear mapping of image features. k is the dimension of the key vector, and softmax is used to calculate the weight of the cross-attention mechanism. Through the above attention mechanism, the prediction model can fully explore the coupling characteristics between power features and image features and reduce the redundancy of feature information.

[0079] In step 4, the present invention uses the Transformer model to further explore the temporal characteristics in the fusion features to obtain deep multimodal temporal features. The specific process can be expressed as follows:

[0080]

[0081] Among them, F final is the multimodal time series feature, f Transformer is the Transformer model, is the multimodal fusion feature at each historical moment, and n is the dimension of the fusion feature; Figure 4 As shown in Figure 1, multi-head attention is the core of the Transformer model, which enables the model to focus on different parts of the input sequence at the same time, thereby improving its attention ability and capturing the complex dependencies between input and output. The mathematical representation of the multi-head attention mechanism is as follows:

[0082] MultiHead(Q,K,V)=Concat(head 1 ,...,head h )W O ,head i =Attention(QW i Q ,KW i K ,VW i V )

[0083] Among them, MultiHead represents the multi-head attention mechanism, Q, K, and V are query vectors, key vectors, and value vectors respectively, Concat represents the concatenation operation, and head i ——The weight result generated by the i-th attention operation, h is the number of implementations of the attention mechanism, W l Q , W l K , W l V , W O is the weight coefficient of the linear mapping.

[0084] In step 5, the present invention obtains the photovoltaic power prediction result y through a multi-layer perceptron based on the power time series characteristics, cloud image spatiotemporal characteristics and deep time series characteristics:

[0085] y=f MLP ([F power ,F image ,F final ]);

[0086] Among them, F MLP It is a multi-layer perceptron model.

[0087] In order to verify the performance of the method proposed in the present invention, the following experiments were conducted in the embodiment. The data set used was and the model performance was evaluated by quantitative error indicators. The satellite cloud images and photovoltaic power data used in this study came from a photovoltaic power station in China (2017-2018). 80% of the data set was used for training and 20% for testing. Three error analysis indicators were used in this embodiment: mean error (MAE), root mean square error (RMSE), and coefficient of determination (R 2 ), defined as follows:

[0088]

[0089]

[0090] Where N represents the total number of test samples, y i and Represent the predicted value and actual value of the i-th sample respectively, Represents the mean of the actual values.

[0091] In order to intuitively reflect the prediction effect, Figure 5The photovoltaic power generation power prediction curve obtained by the method of the present invention 1 hour in advance is shown. It can be seen that the method of the present invention can achieve accurate prediction of photovoltaic power. In order to further evaluate the prediction performance of the prediction model, a total of four comparison models are set, which can be represented as S1, S2, S3, and S4. S1 uses 2DCNN to extract the spatial features of the cloud map, S2 uses ConvLSTM to extract the spatiotemporal features of the cloud map, and S3 uses MLP and 3DCNN to extract the feature information of historical power and satellite cloud maps respectively. In addition, Table 1 shows the prediction error of each model for photovoltaic power prediction 10 minutes in advance. It can be seen that compared with the comparison model, the prediction result of the model constructed based on the present invention has a smaller error and better performance.

[0092] Table 1 Comparison of prediction results of different models

[0093]

[0094] In summary, the photovoltaic power prediction method based on multi-stage time series feature mining designed by the present invention first extracts features of historical photovoltaic power sequences through the LSTM model to obtain power time series features; extracts features of historical satellite cloud image sequences through the 3DCNN model to obtain cloud spatiotemporal dynamic features; based on the residual cross-attention mechanism, the power time series features and cloud spatiotemporal dynamic features are fused to reduce the redundancy of the fused features, improve the availability of feature information, and obtain multimodal fusion features; the Transformer model is used to further mine the time series characteristics in the fused features to obtain deep multimodal time series features; experimental results show that the present invention can improve the prediction accuracy of photovoltaic power, provide support for the safe and stable operation of the power grid, and has certain engineering practical value.

[0095] Based on the same technical concept as the method embodiment, the present invention also provides a photovoltaic power prediction system based on multi-stage time series feature mining, including a power feature extraction module, an image feature extraction module, a feature fusion module, a deep feature extraction module and a time series prediction module, which is used to execute the above-mentioned photovoltaic power prediction method based on multi-stage time series feature mining.

[0096] The power feature extraction module extracts features from the historical photovoltaic power sequence through the LSTM model to obtain power time series features;

[0097] The image feature extraction module extracts features from historical satellite cloud image sequences through a 3DCNN model to obtain the spatiotemporal dynamic characteristics of the cloud layer;

[0098] The feature fusion module fuses the power time series features and the cloud spatiotemporal dynamic features based on the residual cross-attention mechanism to obtain multimodal fusion features;

[0099] The deep feature extraction module uses the Transformer model to further explore the temporal characteristics in the fusion features and obtain deep multimodal temporal features;

[0100] The time series prediction module takes the significant spatiotemporal features of the image and the historical photovoltaic power as input, and outputs the photovoltaic power results for the next 1 to 4 hours through a multi-layer perceptron.

[0101] It should be understood that the photovoltaic power prediction system based on multi-stage time series feature mining in the embodiment of the present invention can implement all the technical solutions in the above-mentioned method embodiment, and the functions of its various functional modules can be specifically implemented according to the methods in the above-mentioned method embodiments. The specific implementation process can refer to the relevant description in the above-mentioned embodiments, which will not be repeated here.

[0102] The present invention also provides a computer device, comprising: one or more processors; a memory; and one or more programs, wherein the one or more programs are stored in the memory and are configured to be executed by the one or more processors, and when the programs are executed by the processors, the steps of the photovoltaic power prediction method based on multi-stage time series feature mining as described above are implemented.

[0103] The present invention also provides a computer-readable storage medium having a computer program stored thereon, and when the computer program is executed by a processor, the steps of the photovoltaic power prediction method based on multi-stage time series feature mining as described above are implemented.

[0104] It will be appreciated by those skilled in the art that embodiments of the present invention may be provided as methods, devices (systems), computer equipment or computer program products. Therefore, the present invention may take the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware. Moreover, the present invention may take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program codes.

[0105] The present invention is described with reference to a flowchart of a method according to an embodiment of the present invention. It should be understood that each process in the flowchart and a combination of processes in the flowchart can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor or other programmable data processing device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the process in the flowchart. Figure 1 A device that specifies functions in a process or multiple processes.

[0106] These computer program instructions may also be stored in a computer-readable memory capable of directing a computer or other programmable data processing device to operate in a specific manner, so that the instructions stored in the computer-readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 A function specified in a process or multiple processes.

[0107] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operating steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing instructions for implementing the process in the computer or other programmable device. Figure 1 The steps of a specified function in a process or multiple processes.

[0108] Finally, it should be noted that the above is only a preferred embodiment of the present invention and is not intended to limit the present invention. Although the present invention has been described in detail with reference to the aforementioned embodiments, it is still possible for those skilled in the art to modify the technical solutions described in the aforementioned embodiments or to make equivalent substitutions for some of the technical features therein. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the protection scope of the present invention.

Claims

1. A photovoltaic power prediction method based on multi-stage time series feature mining, characterized in that: The following steps are involved: Step 1: Extract features of historical photovoltaic power sequences through the LSTM model to obtain power time series features. Extract features of historical photovoltaic power sequences through the LSTM model to obtain power time series features. The process is expressed as follows: Among them, F power is the power timing characteristic, f LSTM is an LSTM model, m is the number of LSTM layers used, [p1,p2,...,p n ] is the historical photovoltaic power series, n is the length of the time series; Step 2: Use the 3DCNN model to extract features from the historical satellite cloud image sequence to obtain the spatiotemporal dynamic characteristics of the cloud layer. The process is expressed as: image =f 3DCNN ([i1,i2,...,i n ]); Among them, F image is the power timing characteristic, f 3DCNN For the 3DCNN model, [i1,i2,...,i n ] is a historical satellite cloud image sequence; 3DCNN is expressed by the following formula: h (l) =f (l) (w (l) *h (l-1) +b (l) ); Step 3: Based on the residual cross-attention mechanism, the power time series features and the cloud spatiotemporal dynamic features are fused to obtain multimodal fusion features. The specific process is as follows: Step 4: Use the Transformer model to further explore the temporal characteristics in the fusion features and obtain deep multimodal temporal features. The specific process is as follows: Step 5: The prediction results are output through the multi-layer perceptron to predict the photovoltaic power in the next 1 to 4 hours. Based on the power time series characteristics, cloud image spatiotemporal characteristics and deep time series characteristics, as well as the descriptions of the above steps 2 and 3, the power characteristics, image characteristics and fusion characteristics are obtained from the historical data. Finally, the prediction result y of the photovoltaic power is obtained by re-learning these three characteristics: y=f MLP ([F power ,F image ,F final ]); Among them, F MLP It is a multi-layer perceptron model.

2. The photovoltaic power prediction method based on multi-stage time series feature mining according to claim 1 is characterized in that: In step 1, LSTM is expressed by the following formula: i t =σ(W xi *x t +W hi *h t-1 +b i ) f t =σ(W xf *x t +W hf *h t-1 +b f ) g t =σ(W xg *x t +W hg *h t-1 +b g ) Among them, x t ,h t , c t They represent the input value, output value, and unit state of LSTM at time t respectively; h t-1 is the output value of LSTM at time t-1, i t , f t , g t They are input gate, forget gate and output gate, which are used to control the flow of information; W xi , W hi , b i , W xf , W hf , b f , W xc , W hc , b c , W xg , W hg , b g All represent learnable parameters, tanh is the hyperbolic tangent function, σ(·) is the activation function, "*" is the ordinary matrix product, Represents the Hadamard product.

3. The photovoltaic power prediction method based on multi-stage time series feature mining according to claim 1 is characterized in that: In step 2, * represents the convolution operation, w (l) is the weight of the lth convolutional layer, b (l) is the bias of the lth convolutional layer, f (l) is the activation function of the lth convolutional layer, h (l) ,h (l-1) are the outputs of the lth and l-1th convolutional layers respectively; the process is expressed as: Among them, Y is the output feature map, (x, y, z) is the position index in the output feature map, g is the input feature map of the pooling operation, M, N, O represent the size of the pooling window in height, depth and width respectively, S d , S h , S w are the moving steps of the pooling window in depth, height and width, respectively, and m, n, o are the index offsets corresponding to the current position of the pooling window.

4. The photovoltaic power prediction method based on multi-stage time series feature mining according to claim 1 is characterized in that: Step 3, F fusion is a multimodal fusion feature, RS-Attention is a residual cross attention mechanism, Q, K, and V are query vectors, key vectors, and value vectors, respectively. Q is obtained through nonlinear mapping of power features, and K and V are obtained through nonlinear mapping of image features. k is the dimension of the key vector, and softmax is used to calculate the cross attention mechanism weights.

5. The photovoltaic power prediction method based on multi-stage time series feature mining according to claim 1 is characterized in that: Step 4, F final is the multimodal time series feature, f Transformer is the Transformer model, is the multimodal fusion feature at each historical moment, and n is the dimension of the fusion feature.

6. A photovoltaic power prediction system based on multi-stage time series feature mining, characterized in that: It includes a power feature extraction module, an image feature extraction module, a feature fusion module, a deep feature extraction module and a time series prediction module, and is used to execute the photovoltaic power prediction method based on multi-stage time series feature mining as described in any one of claims 1-5.

7. A photovoltaic power prediction system based on multi-stage time series feature mining according to claim 6, characterized in that: Power feature extraction model: extract features of historical photovoltaic power sequences through the LSTM model to obtain power time series features; The image feature extraction module extracts features from historical satellite cloud image sequences through a 3DCNN model to obtain the spatiotemporal dynamic characteristics of the cloud layer; The feature fusion module fuses the power time series features and the cloud spatiotemporal dynamic features based on the residual cross-attention mechanism to obtain multimodal fusion features; The deep feature extraction module uses the Transformer model to further explore the temporal characteristics in the fusion features and obtain deep multimodal temporal features; The time series prediction module takes the significant spatiotemporal features of the image and the historical photovoltaic power as input, and outputs the photovoltaic power results for the next 1 to 4 hours through a multi-layer perceptron.

8. A computer device, characterized in that: include: one or more processors; Memory; And one or more programs, wherein the one or more programs are stored in the memory and are configured to be executed by the one or more processors, and when the programs are executed by the processors, the steps of the photovoltaic power prediction method based on multi-stage time series feature mining as described in any one of claims 1-5 are implemented.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the photovoltaic power prediction method based on multi-stage time series feature mining as described in any one of claims 1 to 5 are implemented.

Images