A solar power generation power prediction method and system

Abstract

The application discloses a solar power generation power prediction method and system, and relates to the technical field of solar power generation. The solar power generation power prediction method obtains high-frequency time sequence data of a region where a prediction target is located under a current extreme weather event; and inputs the high-frequency time sequence data into a pre-trained double-channel feature extraction network for parallel processing, and synchronously outputs a power change feature vector and a prediction confidence feature vector of the region where the prediction target is located; based on a gated attention algorithm, the power change feature vector and the prediction confidence feature vector are subjected to weighted decision and regression processing, and a solar power generation power prediction value of the region where the prediction target is located in a set future period is generated. Through the design of the double-channel feature extraction network and the gated attention fusion mechanism, the application realizes accurate and reliable prediction of solar power generation power under extreme weather, and effectively improves the dispatching capability and operation safety level of the power grid under complex meteorological conditions.

Description

Technical Field

[0001] This invention relates to the field of solar power generation technology, specifically to a method and system for predicting solar power generation output. Background Technology

[0002] As the global energy structure shifts towards cleaner and lower-carbon energy, the installed capacity and grid connection rate of solar power generation continue to grow rapidly. Solar power output directly depends on the solar irradiance received by the Earth's surface, exhibiting significant intermittency and volatility, which poses a severe challenge to the real-time balance, dispatching, operation, and safety and stability of the power system. This challenge becomes particularly prominent against the backdrop of frequent extreme weather events.

[0003] Extreme weather events, such as severe convective storms, persistent rain, dust storms, and rapidly passing fragmented clouds, can cause drastic and nonlinear changes in surface irradiance within a short period. These changes result in complex fluctuations in the solar power generation curve, including sharp rises and falls, significantly deviating from the smooth profile observed under clear weather conditions. High-precision and highly reliable power forecasting is crucial for power grid dispatch centers to plan reserve capacity in advance, optimize operational modes, and mitigate safety risks. Therefore, developing technologies that can effectively adapt to and accurately predict changes in solar power generation under extreme weather conditions is of urgent practical significance and important application value for improving the absorption capacity, operational resilience, and power supply security of grids with a high proportion of renewable energy.

[0004] Existing technologies, such as the invention patent application CN119944677B, disclose a method, apparatus, electronic device, and storage medium for predicting solar power generation. The method includes: predicting a first power generation of a solar power generation device using a physical model based on physical parameters of a target area; the physical model is a mapping model between physical parameters and power generation constructed based on the laws of solar celestial motion and photovoltaic power generation mechanisms; predicting a second power generation of the solar power generation device using a deep learning model based on meteorological data of the target area and historical power generation data of the solar power generation device; and fusing the first and second power generations based on the weather type of the target area to obtain a third power generation of the solar power generation device. The solar power generation prediction method, apparatus, electronic device, and storage medium provided in this disclosure can improve the reliability of solar power generation prediction, facilitating accurate grid dispatching and energy management.

[0005] Based on the above-mentioned solutions, the limitations of existing technologies include at least the following problems: In the construction of power prediction models, they are limited to using pre-trained fixed models to perform macroscopic classification and result fusion for different weather types, without specifically modeling and real-time analysis of the rapidly changing and drastically fluctuating dynamic characteristics of extreme weather events. This may result in the prediction performance meeting requirements under normal weather conditions, but when dealing with extreme scenarios such as sudden changes in irradiance and rapid cloud migration, there is a serious mismatch between the model's dynamic response capability and prediction accuracy. This leads to a series of negative impacts such as prediction lag, error amplification, and even failure, which greatly inhibits the practical value of the prediction system in emergency grid dispatch and safety early warning, and is not conducive to providing reliable guarantees for the safe, stable, and efficient operation of high-proportion renewable energy grids under extreme weather conditions. Summary of the Invention

[0006] To address the shortcomings of existing technologies, this invention provides a method and system for predicting solar power generation, which solves the problems of decreased prediction accuracy, response lag, and reduced reliability caused by insufficient dynamic adaptability of models and rigid feature extraction and fusion mechanisms when dealing with extreme weather events.

[0007] To achieve the above objectives, the present invention provides the following technical solution: a method for predicting solar power generation, comprising the following steps: acquiring high-frequency time-series data of the area where the target is located under the current extreme weather event, wherein the high-frequency time-series data consists of total irradiance values, direct radiation values, and diffuse radiation values ​​at several time points; The high-frequency time-series data is input into a pre-trained dual-channel feature extraction network for parallel processing, and the power change feature vector and prediction confidence feature vector of the target area are output synchronously. The pre-trained dual-channel feature extraction network is specifically composed of a power change feature extraction channel and a prediction confidence feature extraction channel. Based on the gated attention algorithm, the power change feature vector and the prediction confidence feature vector are subjected to weighted decision and regression processing to generate the predicted solar power generation power of the target area in a set future time period.

[0008] Furthermore, the power change feature extraction channel is composed of a one-dimensional convolutional layer and a gated recurrent unit network layer connected in sequence.

[0009] Furthermore, the specific steps for outputting the power change feature vector of the predicted target area are as follows: In the one-dimensional convolutional layer of the power change feature extraction channel, the high-frequency time series data is convolved to extract its local features; in the gated recurrent unit network layer of the power change feature extraction channel, the local features are processed by time series modeling, and its final state is output as a power change feature vector.

[0010] Further, the specific steps for performing convolution processing on the high-frequency time-series data to extract its local features are as follows: using multiple convolution kernels of different sizes, one-dimensional convolution operations are performed on the high-frequency time-series data respectively; the operation results of each size convolution kernel are concatenated along the channel dimension, and the concatenated feature tensor is output as a local feature.

[0011] Further, the specific steps for performing time-series modeling on the local features and outputting their final state as a power change feature vector are as follows: the local features are used as the input sequence; the hidden state of the current time step is calculated based on the input of each time step and the hidden state of the previous time step through the update gate and reset gate of the gated recurrent unit network layer; the hidden state calculated in the last time step is output as the power change feature vector.

[0012] Furthermore, the prediction confidence feature extraction channel is composed of a one-dimensional convolutional layer and a multi-head self-attention network layer connected in sequence.

[0013] Furthermore, the specific steps for outputting the power change feature vector of the predicted target region are as follows: In the one-dimensional convolutional layer of the prediction confidence feature extraction channel, convolution operations and feature transformations are performed on the high-frequency time series data to output an intermediate feature sequence; In the multi-head self-attention network layer of the prediction confidence feature extraction channel, self-attention calculation is performed on the intermediate feature sequence to evaluate the correlation between features at each time step within the sequence and aggregate global information to output the prediction confidence feature vector.

[0014] Further, the specific steps for performing convolution operations and feature transformations on the high-frequency time-series data to output an intermediate feature sequence are as follows: performing convolution operations on the high-frequency time-series data using multiple one-dimensional convolution kernels, and performing nonlinear activation through a linear rectified function to obtain a primary feature map; transposing and reshaping the primary feature map in terms of dimension to obtain an intermediate feature sequence.

[0015] Further, the specific steps for performing self-attention calculation on the intermediate feature sequence, evaluating the correlation between features at each time step within the sequence, and aggregating global information are as follows: The intermediate feature sequence is sequentially transformed linearly to generate corresponding query vector matrices, key vector matrices, and value vector matrices; in each independent attention head, the dot product of the query vector matrix and the key vector matrix is ​​calculated, scaled, and then the Softmax function is applied to obtain the attention weight matrix; the attention weight matrix is ​​used to perform a weighted summation on the value vector matrix to obtain the output of that attention head; the outputs of all attention heads are concatenated along the feature dimension, and the concatenated result is aggregated into a prediction confidence feature vector through a linear projection.

[0016] A solar power generation prediction system includes: a multi-source high-frequency data acquisition unit for acquiring high-frequency time-series data of the target area under current extreme weather events, wherein the high-frequency time-series data consists of total irradiance values, direct radiation values, and diffuse radiation values ​​at several time points; a dual-channel feature generation unit for inputting the high-frequency time-series data into a pre-trained dual-channel feature extraction network for parallel processing, and simultaneously outputting a power change feature vector and a prediction confidence feature vector of the target area, wherein the pre-trained dual-channel feature extraction network specifically consists of a power change feature extraction channel and a prediction confidence feature extraction channel; and a gated attention regression unit for performing weighted decision and regression processing on the power change feature vector and the prediction confidence feature vector based on a gated attention algorithm to generate a predicted solar power generation value for the target area in a set future time period.

[0017] The present invention has the following beneficial effects: (1) The solar power generation prediction method uses parallel processing of power change feature extraction channel and prediction confidence feature extraction channel, so that the model can not only accurately extract power change features, but also evaluate the prediction confidence during the prediction process. This dual-channel structure can fully explore the multi-dimensional information in time series data. Especially when facing extreme weather events, it can enhance the model's stability and prediction accuracy. Compared with the traditional single feature extraction method, the present invention can better capture the changing trend of solar power generation and adjust the prediction strategy according to real-time data, thereby achieving more accurate power prediction. Especially in extreme weather or complex environments, it can provide more reliable data support for power management departments.

[0018] (2) The solar power generation prediction method improves the weighted decision and regression processing of the power change feature vector and prediction confidence feature vector by adopting the gated attention algorithm. Traditional methods often rely on fixed weights to integrate different features when processing time series data. However, the gated attention algorithm in this invention can dynamically assign different weights to the input of each time step and pay more attention to key features. This adaptive weighted decision mechanism significantly improves the prediction effect of the model. Especially when facing complex and uncertain environmental changes, it can optimize the results by adjusting the weights. With the help of the gated attention mechanism, the model can not only strengthen the response to key features, but also adjust itself in the prediction process, so as to make the power prediction more accurate and the results more stable.

[0019] (3) This solar power generation prediction method uses a variety of convolution kernels of different sizes to perform one-dimensional convolution operations on high-frequency time series data, which fully explores the local features and overall information in the time series data. With this approach, the model can simultaneously focus on features at different scales, thereby improving its ability to handle diverse and complex data situations. Especially when dealing with complex factors such as weather fluctuations, the diverse settings of convolution kernels enable the model to capture subtle changes, which is crucial for predicting the trend of solar power generation. At the same time, the gated recurrent unit network layer provides stronger temporal correlation learning capabilities in temporal modeling, enabling features to be effectively extracted not only in space but also reliably modeled in the time dimension, thus making the prediction results more accurate and reliable. The introduction of multi-head self-attention network enables the model to extract effective features from the overall information and further optimize feature extraction and fusion by combining the correlation between different time steps.

[0020] (4) The solar power generation prediction system significantly improves the system's response speed and prediction accuracy by adopting a multi-source high-frequency data acquisition unit and a dual-path feature generation unit. The multi-source high-frequency data acquisition unit can accurately capture the irradiance changes in the solar power generation area under extreme weather events, thereby providing comprehensive and timely data support for subsequent predictions. This data acquisition method ensures that the system can make predictions based on the latest weather changes, avoiding the data delay problem in traditional methods. The dual-path feature generation unit processes high-frequency time-series data in parallel and extracts power change features and prediction confidence features respectively. This not only ensures efficient data processing but also evaluates the credibility of the results in real-time prediction, improving the system's ability to cope with complex environments. With the help of the weighted decision-making and regression processing of the gated attention regression unit, the system can generate solar power generation prediction values ​​more accurately, thereby providing scientific and reliable decision-making basis for power management departments.

[0021] Of course, any product implementing this invention does not necessarily need to achieve all of the advantages described above at the same time. Attached Figure Description

[0022] Figure 1 This is a flowchart of a solar power generation prediction method according to the present invention.

[0023] Figure 2 This is a flowchart illustrating the specific steps involved in outputting the power change feature vector of the target region in a solar power generation prediction method according to the present invention.

[0024] Figure 3 This is a block diagram of a solar power generation prediction system according to the present invention. Detailed Implementation

[0025] Please see Figure 1This invention provides a technical solution: a method for predicting solar power generation, comprising the following steps: acquiring high-frequency time-series data of the area where the prediction target (i.e., solar power generation equipment) is located under the current extreme weather event, wherein the high-frequency time-series data consists of total irradiance values, direct radiation values, and diffuse radiation values ​​at several time points; inputting the high-frequency time-series data into a pre-trained dual-channel feature extraction network for parallel processing, and simultaneously outputting a power change feature vector and a prediction confidence feature vector of the area where the prediction target is located, wherein the pre-trained dual-channel feature extraction network specifically consists of a power change feature extraction channel and a prediction confidence feature extraction channel; and performing weighted decision and regression processing on the power change feature vector and the prediction confidence feature vector based on a gated attention algorithm to generate a predicted value of solar power generation in the area where the prediction target is located in a set future period.

[0026] Among them, the power change feature vector generated based on all the aforementioned steps (denoted as ) and the predicted confidence feature vector (denoted as The specific steps for generating the predicted power value through weighted decision-making and regression processing using a gated attention algorithm are as follows: Feature splicing and gating signal generation: Feature concatenation: concatenates two feature vectors into a joint feature vector, i.e.: ; Generate gated weights: Pass the concatenated vector through a fully connected layer (or gated weight generator) and then through a sigmoid activation function to generate a scalar gated weight. This weight is used to dynamically balance the importance of the two eigenvectors, that is: ; in, and These are trainable parameters. It's the Sigmoid function, outputting... ; Gated attention-weighted fusion: Feature transformation: Mapping the two original feature vectors through independent linear layers to the same fusion space dimension. : ; ; in, , , These are trainable parameters; Gated fusion: using the generated gating weights The two transformed features are weighted and summed to obtain the fused feature vector: ; The core of this step is the attention mechanism: The size is determined by the joint features The decision is essentially an adaptive attention score that assesses whether power change features or prediction confidence features should be given more weight given the current specific input (extreme weather data). Regression Prediction: fuse feature vectors A regressor (usually a multi-layer fully connected network) is used to directly map to the final target, namely the solar power generation prediction sequence for a future time period (e.g., the next 24 hours, with 96 points at 15-minute intervals). Regression calculation: ; in, , These are the parameters of the first fully connected layer; It is a non-linear activation function; , These are the parameters of the second (output) fully connected layer; T is the number of predicted points for a given future time period (e.g., 96). This is the generated sequence of predicted solar power output values; Output: Vector Each element This corresponds to the predicted power value for the t-th time interval in the future. These values ​​can be directly used in grid dispatching or energy management systems.

[0027] Specifically, the power change feature extraction channel is composed of a one-dimensional convolutional layer and a gated recurrent unit network layer connected in sequence.

[0028] like Figure 2 As shown, the specific steps for outputting the power change feature vector of the predicted target area are as follows: In the one-dimensional convolutional layer of the power change feature extraction channel, high-frequency time series data is convolved to extract its local features; in the gated recurrent unit network layer of the power change feature extraction channel, the local features are processed by time series modeling, and its final state is output as the power change feature vector.

[0029] The specific steps for performing convolution processing on high-frequency time-series data to extract its local features are as follows: Using convolution kernels of various sizes, one-dimensional convolution operations are performed on high-frequency time-series data, specifically as follows: High-frequency time series data are constructed into a two-dimensional tensor. ; Where L represents the total number of time steps (sequence length). This represents the input feature dimension, which is 3 here (i.e., total irradiance, direct radiation, and diffuse radiation values). Let the predefined set of convolution kernel sizes be... ,For example: , , Its unit is time point, which is used to capture local patterns at short, medium and long time scales respectively; For the k-th convolutional kernel, its parameter is a three-dimensional weight tensor. and a bias vector ; in, This refers to the number of feature map channels output by a convolutional kernel of this size. A common setting is to set all... equal; At each output position i (where , Assuming "effective" padding is used and the stride is 1), the convolution operation is essentially a local window of the convolution kernel on the input sequence. The weighted summation and bias are applied, followed by a nonlinear activation, as shown in the following formula: ; in, This represents a nonlinear activation function. In this embodiment, the ReLU (Rectified Linear Unit) function is used, which is defined as... This function introduces nonlinearity and helps alleviate the gradient vanishing problem. Performing this operation on all output positions i yields the complete output feature map corresponding to the k-th convolutional kernel. ; The results of convolutional kernels of different sizes are concatenated along the channel dimension, and the concatenated feature tensor is output as a local feature. Specifically: Obtain feature maps of all K convolutional kernel sizes Then, they are concatenated along the second dimension (i.e., the channel dimension). The specific operation is as follows: For each identical time position i, all feature vectors output by different convolutional kernels at that position (i.e., ...) Connect the first and last parts together to form a longer feature vector; In mathematics, the concatenated feature tensor Total number of channels Each row of the tensor F (corresponding to a time step) incorporates local abstract representations of the original data at different time scales, and is therefore defined as a local feature, serving as input for the subsequent time series modeling module; The specific steps for performing time-series modeling on local features and outputting their final state as a power change feature vector are as follows: Using local features as the input sequence, specifically: The obtained local feature tensor Conceptually reinterpreted as a sequence of data; The length of the sequence is The value of the t-th element (time step t) in the sequence is the t-th row of F, i.e. ; Therefore, the entire local feature tensor is transformed into an input sequence. The data is prepared to be input into the recurrent neural network layer for time series modeling. By using the update and reset gates of the gated recurrent unit network layer, the hidden state of the current time step is calculated based on the input of each time step and the hidden state of the previous time step. Specifically: Let the hidden state dimension of the gated recurrent unit network layer be . ; Initialize the hidden state of the first time step The zero vector is usually the zero vector. ; The core of GRU lies in its two gating mechanisms: the UpdateGate and the ResetGate, for time step t (from 1 to 2). The calculation process is as follows: Computational update gate The update gate determines how much past information is available. It needs to be retained in the current state. It is the previous hidden state. With current input The result of a linear transformation of the concatenated vectors, followed by the Sigmoid activation function, is: ; in, This represents vector concatenation. and These are trainable parameters. The output is compressed to the (0,1) interval to simulate the opening of the "gate". Calculate the reset door The reset gate determines how much past information needs to be "forgotten" to help the model learn short-term dependencies that are more relevant to the current input. Its calculation is similar to the update gate, but it uses different parameters: ; Calculate candidate hidden state The candidate state is a potential new state calculated based on the current input and the previous state after being "filtered" by the reset gate. It uses the hyperbolic tangent (tanh) activation function, i.e.: ; in, This represents element-wise multiplication (Hadamard product). This indicates that some past information is being selectively retained. Calculate the current hidden state The final hidden state is the linear interpolation of the previous state and the candidate state under update gate control, that is: ; This formula enables dynamic control of information flow: when When the value is close to 0, the model primarily retains past memories; when When the value is close to 1, more new information from the current candidate will be adopted; The hidden state calculated in the last time step is output as the power change feature vector, specifically as follows: The above recursive calculation is performed to the last time step of the sequence. To obtain the final hidden state. This vector encodes the contextual information and long-term dependencies of the entire input sequence (i.e., the local feature sequence) and is defined as the final output of the power change feature extraction channel, i.e., the power change feature vector. This vector will be fed into the subsequent fusion module to generate the final prediction.

[0030] In this implementation scheme, the combination of one-dimensional convolutional layers and gated recurrent unit (GRU) network layers effectively extracts local features and temporal dependencies from high-frequency time-series data. Using convolutional kernels of different sizes to process the data enables the model to capture multi-dimensional features across short, medium, and long time scales, fully reflecting the multi-layered characteristics of the data. This method, combining convolution with temporal modeling, enhances the processing capability for complex time-series data, especially in the face of extreme weather events, more accurately reflecting fluctuations in solar power generation. Through the update and reset gate mechanisms of the GRU network, the model can flexibly control information flow, deciding which historical data should be retained and which should be ignored, thus ensuring the effective transmission of important information over long time series and avoiding information loss or over-reliance on short-term data. This structure not only improves the model's robustness in processing complex time-series data but also enhances its adaptability under different prediction scenarios, making the prediction of solar power generation more accurate and reliable. The final output power change feature vector provides accurate input for subsequent predictions.

[0031] Specifically, the prediction confidence feature extraction channel is composed of a one-dimensional convolutional layer and a multi-head self-attention network layer connected in sequence.

[0032] The specific steps for outputting the power change feature vector of the predicted target region are as follows: In the one-dimensional convolutional layer of the prediction confidence feature extraction channel, convolution operations and feature transformations are performed on the high-frequency time series data to output the intermediate feature sequence; in the multi-head self-attention network layer of the prediction confidence feature extraction channel, self-attention calculation is performed on the intermediate feature sequence to evaluate the correlation between features at each time step within the sequence and aggregate global information to output the prediction confidence feature vector.

[0033] The specific steps for performing convolution operations and feature transformations on high-frequency time-series data to output intermediate feature sequences are as follows: Multiple one-dimensional convolutional kernels are used to perform convolution operations on high-frequency time-series data, and non-linear activation is performed through a linear rectified function to obtain a primary feature map, which is as follows: Let the input high-frequency time-series data be... .use One-dimensional convolution is performed using 1 convolution kernels, with each kernel having a weight of . , bias is ; For the c-th convolutional kernel, the convolution output at position c is: ; After computation for all positions i and all convolution kernels c, a linear output is obtained. ,in Then, the ReLU activation function is applied to each element of Z: ; Obtain the primary feature map ; The primary feature map is transposed and reshaped in dimensionality to serve as an intermediate feature sequence, specifically as follows: The primary feature map P is considered as a two-dimensional matrix. To transform it into a sequence format suitable for self-attention networks, the following steps are performed: First, transpose P to obtain ; Then, the transposed matrix is ​​redimensionalized and reinterpreted as a sequence; Specifically Each column is considered as a feature vector of a time step in the sequence. Therefore, the intermediate feature sequence S can be represented as: ,in ; Alternatively, in implementation, we can directly let S=P, and treat each row of S as a feature of a time step (i.e. ); Finally, the intermediate feature sequence ; in, For sequence length, The feature dimension for each time step; The specific steps for performing self-attention computation on the intermediate feature sequence, evaluating the correlation between features at different time steps within the sequence, and aggregating global information are as follows: The intermediate feature sequences are sequentially transformed using linear transformations to generate the corresponding query vector matrix, key vector matrix, and value vector matrix, as follows: Let the intermediate feature sequence be ,in ; Introduce three trainable parameter matrices: Query weight matrix ; Key weight matrix ; Value weight matrix ; The query matrix Q and the key matrix are generated through the following linear transformation. K and sum matrix V: ; ; ; in, , , and These are the dimensions for keys / queries and values, respectively. In each independent attention head, the dot product of the query vector matrix and the key vector matrix is ​​calculated, scaled, and then the Softmax function is applied to obtain the attention weight matrix, which is as follows: For an attention head, compute the scaled dot product attention; First, calculate the dot product of Q and K, then divide by... Scaling is applied to prevent the gradient from becoming too large or too small: ; The resulting score matrix The element in the i-th row and j-th column represents the degree of attention given by the i-th time step to the j-th time step in the sequence. Then, the Softmax function is applied to each row for normalization, so that the sum of all weights in each row is 1, i.e.: ; in, ; The obtained matrix This is the attention weight matrix; The value vector matrix is ​​weighted and summed using the attention weight matrix to obtain the output of the attention head, which is as follows: Multiplying the attention weight matrix A by the value matrix V yields the output matrix of the attention head. : ; in, The essence of this operation is that, for each time step i in the output sequence, its new feature representation is the feature of all time steps in the original sequence. The weighted sum, with weights determined by Decide; The outputs of all attention heads are concatenated along the feature dimension, and the concatenated result is aggregated into a prediction confidence feature vector through a linear projection, specifically as follows: Assume there are H attention heads in total, each head producing an output. The outputs of all the heads are concatenated along the feature dimension (column direction) to form a concatenated matrix. : ; Then, through a trainable linear projection matrix The concatenated features are mapped to the final output dimension. (generally (consistent with the expected feature vector dimension) ; in, ; Finally, in order to aggregate the information of the entire sequence into a single vector representation, it is usually taken as... The mean or the output of the last time step along the sequence dimension (first dimension). In this embodiment, mean pooling is used: ; The resulting vector This is the feature vector of prediction confidence.

[0034] In this implementation scheme, multiple convolutional kernels are used for convolution operations, which can extract local features from different time scales and improve the ability to capture data details. The ReLU activation function is used to handle nonlinear features, further enhancing the model's expressive power. Through transpose and reshaping operations, the convolutional output is transformed into a sequence format suitable for processing by self-attention networks, enabling the model to handle complex relationships between different time steps. The multi-head self-attention mechanism can evaluate the correlation between features at each time step globally. By using multiple attention heads to focus on different feature dimensions, the weight distribution of information is optimized, thereby improving the model's performance in complex time-series data. After linear projection and mean pooling, the obtained prediction confidence feature vector not only improves the reliability of the prediction results but also enhances the model's adaptability in dynamically changing environments. This scheme can provide a more accurate and stable foundation for solar power generation prediction, and has higher application value and practical effect, especially in dealing with extreme weather changes.

[0035] Please see Figure 3 This invention provides a technical solution: a solar power generation prediction system, comprising: a multi-source high-frequency data acquisition unit, used to acquire high-frequency time-series data of the area where the prediction target is located under the current extreme weather event, the high-frequency time-series data consisting of total irradiance values, direct radiation values, and diffuse radiation values ​​at several time points; a dual-channel feature generation unit, used to input the high-frequency time-series data into a pre-trained dual-channel feature extraction network for parallel processing, and simultaneously output the power change feature vector and prediction confidence feature vector of the area where the prediction target is located, the pre-trained dual-channel feature extraction network specifically consisting of a power change feature extraction channel and a prediction confidence feature extraction channel; and a gated attention regression unit, used to perform weighted decision and regression processing on the power change feature vector and prediction confidence feature vector based on a gated attention algorithm, to generate a predicted value of solar power generation in the area where the prediction target is located in a set future period.

[0036] Although preferred embodiments of the invention have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including both the preferred embodiments and all changes and modifications falling within the scope of the invention.

[0037] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the claims of this invention and their equivalents, this invention also intends to include these modifications and variations.

Claims

1. A method for predicting solar power generation, characterized in that, Includes the following steps: The high-frequency time-series data of the area where the target is located under the current extreme weather event are obtained. The high-frequency time-series data consists of the total irradiance value, direct radiation value, and diffuse radiation value at several time points. The high-frequency time-series data is input into a pre-trained dual-channel feature extraction network for parallel processing, and the power change feature vector and prediction confidence feature vector of the target area are output synchronously. The pre-trained dual-channel feature extraction network is specifically composed of a power change feature extraction channel and a prediction confidence feature extraction channel. Based on the gated attention algorithm, the power change feature vector and the prediction confidence feature vector are subjected to weighted decision and regression processing to generate the predicted solar power generation power of the target area in a set future time period.

2. The solar power generation prediction method according to claim 1, characterized in that, The power change feature extraction channel is composed of a one-dimensional convolutional layer and a gated recurrent unit network layer connected in sequence.

3. The solar power generation prediction method according to claim 2, characterized in that, The specific steps for outputting the power change feature vector of the predicted target region are as follows: In the one-dimensional convolutional layer of the power change feature extraction channel, the high-frequency time-series data is convolved to extract its local features; In the gated recurrent unit network layer of the power change feature extraction channel, the local features are processed by time series modeling, and their final state is output as a power change feature vector.

4. The solar power generation prediction method according to claim 3, characterized in that, The specific steps for performing convolution processing on the high-frequency time-series data to extract its local features are as follows: One-dimensional convolution operations are performed on the high-frequency time-series data using convolution kernels of various sizes. The operation results of convolution kernels of various sizes are concatenated along the channel dimension, and the concatenated feature tensor is output as a local feature.

5. The solar power generation prediction method according to claim 4, characterized in that, The specific steps for performing time-series modeling on the local features and outputting their final state as a power change feature vector are as follows: Use the local features as the input sequence; By using the update gate and reset gate of the gated recurrent unit network layer, the hidden state of the current time step is calculated based on the input of each time step and the hidden state of the previous time step. The hidden state calculated at the last time step is output as the power change feature vector.

6. The solar power generation prediction method according to claim 1, characterized in that, The prediction confidence feature extraction channel is composed of a one-dimensional convolutional layer and a multi-head self-attention network layer connected in sequence.

7. The solar power generation prediction method according to claim 6, characterized in that, The specific steps for outputting the power change feature vector of the predicted target region are as follows: In the one-dimensional convolutional layer of the prediction confidence feature extraction channel, convolution operations and feature transformations are performed on the high-frequency time series data to output an intermediate feature sequence; In the multi-head self-attention network layer of the prediction confidence feature extraction channel, self-attention calculation is performed on the intermediate feature sequence to evaluate the correlation between features at each time step within the sequence and aggregate global information to output the prediction confidence feature vector.

8. The solar power generation prediction method according to claim 7, characterized in that, The specific steps for performing convolution and feature transformation on the high-frequency time-series data to output the intermediate feature sequence are as follows: Multiple one-dimensional convolution kernels are used to perform convolution operations on the high-frequency time-series data, and non-linear activation is performed through a linear rectified function to obtain a primary feature map. The primary feature map is transposed and reshaped in dimensionality to serve as an intermediate feature sequence.

9. The solar power generation prediction method according to claim 8, characterized in that, The specific steps for performing self-attention calculation on the intermediate feature sequence, evaluating the correlation between features at each time step within the sequence, and aggregating global information are as follows: The intermediate feature sequence is sequentially transformed using linear transformations to generate the corresponding query vector matrix, key vector matrix, and value vector matrix; In each independent attention head, the dot product of the query vector matrix and the key vector matrix is ​​calculated, and after scaling, the Softmax function is applied to obtain the attention weight matrix; The value vector matrix is ​​weighted and summed using the attention weight matrix to obtain the output of the attention head. The outputs of all attention heads are concatenated along the feature dimension, and the concatenated result is aggregated into a prediction confidence feature vector through a linear projection.

10. A solar power generation prediction system, employing the solar power generation prediction method according to any one of claims 1-9, characterized in that, include: A multi-source high-frequency data acquisition unit is used to acquire high-frequency time-series data of the area where the predicted target is located under the current extreme weather event. The high-frequency time-series data consists of total irradiance value, direct radiation value, and diffuse radiation value at several time points. The dual-channel feature generation unit is used to input the high-frequency time series data into a pre-trained dual-channel feature extraction network for parallel processing, and simultaneously output the power change feature vector and prediction confidence feature vector of the region where the target is located. The pre-trained dual-channel feature extraction network is specifically composed of a power change feature extraction channel and a prediction confidence feature extraction channel. The gated attention regression unit is used to perform weighted decision and regression processing on the power change feature vector and the prediction confidence feature vector based on the gated attention algorithm, so as to generate the predicted value of solar power generation in the target area in a set future time period.

Images