CNN-BiLSTM-SE-based wide-area photovoltaic station power ultra-short-term prediction method and device

By constructing a photovoltaic site power prediction model based on CNN-BiLSTM-SE, and utilizing the power and geographical information of neighboring sites, the problems of reliance on meteorological data and insufficient accuracy in traditional methods are solved, achieving high-precision photovoltaic power prediction under cloudy weather conditions and supporting grid stability.

CN118920453APending Publication Date: 2024-11-08ZHEJIANG UNIV OF TECH
View PDF 0 Cites 1 Cited by

Patent Information

Application Number
CN202410947604.0
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-07-16
Publication Date
2024-11-08

AI Technical Summary

Technical Problem

Existing methods for ultra-short-term power forecasting of photovoltaic sites rely on expensive meteorological data and have limited forecast accuracy. They are difficult to accurately capture meteorological change trends under cloudy weather conditions, which poses a threat to the stability of the power grid.

Method used

A wide-area photovoltaic (PV) power ultra-short-term prediction method based on CNN-BiLSTM-SE is adopted. By collecting power and geospatial information of multiple PV sites in neighboring areas, dynamic three-dimensional spatiotemporal structure data is constructed. Combining convolutional neural networks, bidirectional long short-term memory networks and attention mechanisms, a PV ultra-short-term power prediction model is built. The model’s attention to different spatiotemporal factors is dynamically adjusted to avoid dependence on meteorological data.

Benefits of technology

It improves the accuracy and reliability of photovoltaic power forecasting, reduces forecasting costs, and can accurately predict changes in photovoltaic site power under cloudy weather conditions, supporting the safe and stable operation of the power system.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118920453A_ABST
    Figure CN118920453A_ABST
Patent Text Reader

Abstract

The invention discloses a CNN-BiLSTM-SE-based wide-area photovoltaic station power ultra-short-term prediction method and device, and the method comprises the steps: original feature data collection and data preprocessing, construction of dynamic three-dimensional space-time structure data 3DData and one-dimensional power structure data Pdata, gradual formation of a historical input-output data set RealCon, and construction of a CNN-BiLSTM-SE prediction model. Setting a model structure parameter and a training hyper-parameter of the photovoltaic prediction model, training the model, loading and starting the prediction model to predict the photovoltaic power, monitoring the performance of the model according to a photovoltaic power prediction error evaluation index, if a monitoring error exceeds a performance threshold sigma, retraining the model, and if not, retraining the model; and continuing to predict the output power of the target photovoltaic station by using the latest photovoltaic prediction model. According to the method, meteorological data is not needed, the space-time dynamic characteristics of a large number of photovoltaic stations in the area can be effectively utilized for photovoltaic station ultra-short-term power prediction, the CNN-BiLSTM-SE prediction model can extract key space-time characteristics in a data set, the photovoltaic power prediction precision is remarkably improved, and the method has good practicability and popularization value.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the field of ultra-short-term prediction of photovoltaic power, and in particular to a method and device for ultra-short-term prediction of wide-area photovoltaic site power based on CNN-BiLSTM-SE. Background Art

[0002] Photovoltaic power generation is a technology that uses light energy to convert photons into electrical energy. It has become an important part of the clean energy field. With the increasing attention paid to environmental protection and sustainable development, photovoltaic power generation systems have been widely used around the world. Photovoltaic power generation has the advantages of environmental protection, renewable, and distributed layout, and has gradually become an important choice to replace traditional fossil energy. In the context of the rapid growth of the current photovoltaic power generation industry, the volatility and intermittent characteristics of photovoltaic power generation systems pose a challenge to the stability of the power grid. The large-scale introduction of photovoltaic power generation systems may have an adverse effect on the operation of the power grid, which may lead to serious economic losses and even threaten the security of the power grid. In order to better manage and optimize the operation of photovoltaic power generation systems and achieve reliable power supply of power systems, it is crucial to accurately predict photovoltaic power. Therefore, accurately predicting photovoltaic power generation has become an important link in ensuring the reliable operation of the power grid.

[0003] However, the ultra-short-term accurate prediction of photovoltaic site power faces many difficulties and challenges: traditional photovoltaic power prediction methods usually rely on meteorological data, such as solar radiation, wind speed and temperature, but these data collection costs are high and are affected by geographical location and weather conditions, and the prediction accuracy is limited; at the same time, the current meteorological prediction model has the problem of insufficient accuracy, and it is difficult to accurately capture the trend of meteorological changes; in addition, there are certain errors in the modeling and parameter setting of distributed photovoltaic sites, which will also lead to a decrease in the accuracy of photovoltaic power prediction based on meteorological forecast information. In view of the difficult prediction scenario of rapid fluctuations in photovoltaic power under cloudy weather, this patent proposes a method for ultra-short-term prediction of wide-area photovoltaic site power that does not rely on meteorological data. By collecting power data and geospatial information of multiple photovoltaic sites in neighboring areas, deep learning technology can be used to more accurately predict the power of each photovoltaic site, reduce prediction costs, and improve prediction accuracy, thereby providing strong support for the safe and stable operation of the power system. Summary of the invention

[0004] In view of the shortcomings of the existing ultra-short-term prediction methods for photovoltaic site power, the present invention proposes a method and device for ultra-short-term prediction of wide-area photovoltaic site power based on CNN-BiLSTM-SE.

[0005] The ultra-short-term prediction method of wide-area photovoltaic site power based on CNN-BiLSTM-SE of the present invention specifically comprises the following steps:

[0006] S1: Original feature data collection and preprocessing: Obtain multi-dimensional real-time data of multiple sites in the wide-area photovoltaic cluster that are independent of meteorological data, and construct an input data matrix. The data matrix of the i-th site at the k-th time is recorded as D ik , then the data matrix D ik Perform outlier and missing value processing and normalization to form a standardized matrix for the target site

[0007] S2: construct dynamic three-dimensional spatiotemporal structure data 3D_Data and one-dimensional power structure data P_data, wherein the standardized data in step S1 are stacked and stored in 3D_Data in the time sequence direction. P_data is stored in stacks in the time sequence direction Normalize the power data of the target site; determine the prediction association step w_size and the prediction duration h, create input-output data with the association step w_size as the window size in the 3D_Data and P_data time series direction, and store the data set as the historical input-output data set Real_Con;

[0008] S3: Integrate convolutional neural network CNN, bidirectional long short-term memory network BiLSTM and attention mechanism SE to build photovoltaic ultra-short-term power prediction model model_y, which includes five modules: input data processing, abstract feature extraction, attention mechanism adjustment, time series modeling and output data processing; the input data processing module is used to receive and process data; the abstract feature extraction module is used to extract abstract spatial features; the attention mechanism adjustment module is used to adjust the model's attention to data, and the attention mechanism module is integrated into the CNN bypass to dynamically adjust the weight of the convolutional neural network feature map to highlight key information; the time series modeling module is used to capture the time series relationship; the output data processing module is used to generate the final photovoltaic power prediction;

[0009] S4: Set the structure parameters and training hyperparameters of the photovoltaic prediction model model_y, where the model structure parameters include the number of layers, convolution kernel size, number of output channels, pooling size, and activation function parameters of the convolutional neural network (CNN), the number of layers, number of hidden units, and random dropout rate parameters of the bidirectional long short-term memory network (BiLSTM), and the number of channels and scaling parameters of the attention mechanism; the training hyperparameters include the optimization algorithm, maximum number of iterations, batch size, initial learning rate, learning rate decay strategy, and gradient threshold;

[0010] S5: Model training: Use the structural parameters and training hyperparameters set in step S4 to train the model. The training process includes iteratively updating the parameters of the photovoltaic prediction model model_y using an optimization algorithm to minimize the loss function. The training data set is the historical input-output data set Real_Con generated in step S2.

[0011] S6: Load the three-dimensional data in the window of duration w_size associated at the last moment in 3D_Data in step S2 as prediction input data, load and start the prediction model obtained in S5 to predict photovoltaic power, output the prediction value of the future duration h, and denormalize it to the photovoltaic power prediction value according to the normalization parameter in step S1;

[0012] S7: Monitor the performance of the model according to the photovoltaic power prediction error evaluation index. If the monitoring error exceeds the performance threshold σ, return to step S5 and use the latest historical data set Real_Con to retrain the model. If not, continue to use the latest photovoltaic prediction model model_y to predict the output power of the target photovoltaic site.

[0013] Preferably, the original feature data construction and data preprocessing in step 1 specifically include:

[0014] S1.1: Obtain the original characteristic data of several sites in the photovoltaic cluster, including power data P and geographic space longitude information x and latitude information y;

[0015] S1.2: Calculate the solar altitude angle h at each station at each moment. The solar altitude angle corresponding to the kth hour of each day in Beijing time is h k , the specific calculation formula is as follows:

[0016]

[0017] h k =sin -1 (sinδsinx+cosδcosxcos((k-12)×15)) (2)

[0018] Where δ is the solar declination angle, n represents the day of the year, and x is the latitude of the station;

[0019] S1.3: Integrate the data of each station at the kth time into the cluster data matrix D k , which includes the power data P of multiple photovoltaic sites, the solar altitude angle h calculated by S2.2, the longitude x and latitude y of the site, and the cluster data matrix D at the kth moment k The expression is as follows:

[0020]

[0021] Where D ik is the multi-dimensional feature information of the ith photovoltaic station at time k, and the power of the ith photovoltaic station at time k is P ik , the solar altitude angle is h ik , longitude and latitude are (x i ,yi ), n is the number of sites in the wide area site cluster;

[0022] S1.4: For the mth target station, all stations perform difference calculations on the longitude and latitude of the target station to obtain longitude difference data Δx and latitude difference data Δy, and construct a data snapshot D' k ;

[0023] Δx i =x i -x m (4)

[0024] Δy i =y i -y m (5)

[0025]

[0026] S1.5: D' k Processing of infinite outliers and missing values ​​in the data: if the power data P of a certain station exceeds M consecutive missing values, the data of all stations in this period will be removed; if the amount of missing data is less than or equal to M, the cubic spline interpolation method is used to fill the missing values ​​to maintain the continuity and smoothness of the data;

[0027] S1.6: Use the maximum value normalization method to calculate D' k Perform data normalization to obtain a standardized matrix The original data uses the difference between the maximum and minimum values ​​of each type of data as the scaling factor, and is mapped to the (0,1) interval through linear transformation. The specific calculation formula is:

[0028]

[0029] Where x is the original data, x min is the minimum value of this type of data, x max The maximum value of this type of data, save the normalization parameter x min and x max .

[0030] Preferably, the specific process of performing performance monitoring on the model according to the photovoltaic power prediction value evaluation index described in step 7 is as follows:

[0031] The performance index of the photovoltaic power prediction value and the actual collected power data is calculated. The performance index uses the normalized root mean square error RMSE and the average percentage error MAPE to monitor the model performance. Taking the rated power of the photovoltaic site as the benchmark, the specific calculation of the performance index is as follows:

[0032]

[0033] Where P oi , P ei Respectively represent predicted power and actual power; P N is the installed capacity of the predicted site; N is the number of valid prediction points during the day, that is, the number of non-zero values.

[0034] The second aspect of the present invention relates to a wide-area photovoltaic site power ultra-short-term prediction device based on CNN-BiLSTM-SE, which is characterized in that it includes a memory and one or more processors, wherein the memory stores executable code, and when the one or more processors execute the executable code, they are used to implement the wide-area photovoltaic site power ultra-short-term prediction method based on CNN-BiLSTM-SE of the present invention.

[0035] A third aspect of the present invention relates to a computer-readable storage medium having a program stored thereon, which, when executed by a processor, implements the ultra-short-term prediction method of wide-area photovoltaic site power based on CNN-BiLSTM-SE of the present invention.

[0036] The present invention does not require meteorological data and can effectively utilize the spatiotemporal dynamic characteristics of a large number of photovoltaic sites in the region to perform ultra-short-term power forecasting of photovoltaic sites. The CNN-BiLSTM-SE prediction model can extract key spatiotemporal features in the data set, significantly improving the accuracy of photovoltaic power prediction. The method has good practicality and promotion value.

[0037] The beneficial effects of the present invention are as follows:

[0038] (1) The present invention designs an innovative data construction method. By integrating key information such as historical power data, solar altitude angle information, and the spatial location of the site, an effective spatiotemporal correlation data set is constructed to reflect the power variation patterns between different photovoltaic sites. This avoids the reliance of traditional methods on expensive meteorological monitoring equipment, thereby reducing the prediction cost and having strong practicality and promotion value.

[0039] (2) The deep learning network CNN-BiLSTM-SE prediction model constructed by the present invention dynamically adjusts the model's attention to different temporal and spatial factors, thereby more effectively capturing the regularity of photovoltaic power changes and improving the accuracy and reliability of the prediction. BRIEF DESCRIPTION OF THE DRAWINGS

[0040] Figure 1 It is a schematic flow chart of the method of the present invention.

[0041] Figure 2 It is a schematic diagram of the location distribution of the wide area site cluster and 32 target sites according to a specific embodiment of the present invention.

[0042] Figure 3It is a schematic diagram of solar altitude angle data of a wide area site cluster according to a specific embodiment of the present invention.

[0043] Figure 4 It is a network structure diagram of the prediction model of the present invention.

[0044] Figure 5a) to Figure 5d) 15-minute prediction effect diagram of a specific embodiment of the present invention, wherein: Figure 5a ) is a diagram showing the 15-minute forecast effect on a sunny day. Figure 5b ) is a diagram of the 15-minute forecast effect of few clouds. Figure 5c ) is a diagram of the cloudy 15-minute forecast effect. Figure 5d ) is a schematic diagram of the 15-minute rainy day forecast effect. DETAILED DESCRIPTION

[0045] The present invention will be described in detail below according to the accompanying drawings and preferred embodiments, and the purpose and effect of the present invention will become more clear. The present invention will be further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention.

[0046] Example 1

[0047] This embodiment takes a cluster of 62 distributed photovoltaic sites in a county-level city in East China as an example, and selects site 32 as the prediction target site. The schematic diagram of the wide-area site cluster and the location distribution of target site 32 is shown in the attached figure. Figure 2 As shown; the specific steps of the ultra-short-term prediction method for wide-area photovoltaic site power that does not rely on meteorological data of the present invention are as follows:

[0048] S1: Original feature data construction and data preprocessing: Obtain multi-dimensional real-time data of multiple sites in the wide-area photovoltaic cluster that do not rely on meteorological data, and construct an input data matrix. The data matrix of the i-th site at the k-th time is recorded as D ik , then the data matrix D ik Perform outlier and missing value processing and normalization to form a standardized matrix for the target site

[0049] S1.1: Obtain the original characteristic data of 62 sites in the photovoltaic cluster, including power data P and geographic longitude information x and latitude information y, with a data resolution of 15 minutes;

[0050] S1.2: Calculate the solar altitude angle h at each station at each moment. The solar altitude angle corresponding to the kth hour of each day is h k The specific calculation formula is as follows. The time series length of the wide-area site cluster in this specific implementation case is 35033, and the solar altitude angle data is distributed in the range of +80° to -80° as shown in the attached figure. Figure 3 Shown:

[0051]

[0052] h k =sin -1 (sinδsinx+cosδcosxcos((k-12)×15)) (2)

[0053] Where δ is the solar declination angle, n represents the day of the year, and x is the latitude of the station;

[0054] S1.3: Integrate the data of each station at the kth time into the cluster data matrix D k , which includes the power data P of multiple photovoltaic sites, the solar altitude angle h calculated by S2.2, the longitude x and latitude y of the site, and the cluster data matrix D at the kth moment k The expression is as follows:

[0055]

[0056] Where D ik is the multi-dimensional feature information of the ith photovoltaic station at time k, and the power of the ith photovoltaic station at time k is P ik , the solar altitude angle is h ik , longitude and latitude are (x i ,y i ), n is the number of sites in the wide area site cluster;

[0057] S1.4: For target station No. 32, all stations perform difference calculation on the longitude and latitude of the target station to obtain longitude difference data Δx and latitude difference data Δy, and construct data snapshot D' k , taking the 48th time at 12 noon as an example, as shown in the following formula, the four columns correspond to the historical photovoltaic power data of the site and the four-dimensional features of the solar altitude angle, longitude difference, and latitude difference:

[0058] Δx i =x i -x m (4)

[0059] Δy i =y i -y m (5)

[0060]

[0061] S1.5: D' kThe infinite outliers and missing values ​​in the data are processed. If the power data P of a certain station is missing for more than 16 consecutive times, the data of all stations in this period will be eliminated; if the amount of missing data is less than or equal to 16, the cubic spline interpolation method is used to fill the missing values ​​to maintain the continuity and smoothness of the data;

[0062] S1.6: Use the maximum value normalization method to calculate D' k Perform data normalization to obtain a standardized matrix The original data uses the difference between the maximum and minimum values ​​of each type of data as the scaling factor, and is mapped to the (0,1) interval through linear transformation. The specific calculation formula is:

[0063]

[0064] Where x is the original data, x min is the minimum value of this type of data, x max The maximum value of this type of data, save the normalization parameter x min and x max .

[0065] S2: construct dynamic three-dimensional spatiotemporal structure data 3D_Data and one-dimensional power structure data P_data, wherein the standardized data in step S1 are stacked and stored in 3D_Data in the time sequence direction. P_data is stored in stacks in the time sequence direction Normalized power data of the target site; Determine the prediction association time of 4 hours and the prediction time of 15 minutes, create input-output data with the association step size as the window size in the 3D_Data and P_data time series direction, and store the data set as the historical input-output data set Real_Con;

[0066] S2.1: Create a three-dimensional matrix 3D_Data with a data structure of "62×35033×4", in which the time series quantity direction can be dynamically expanded, and the standardized data in step S2.6 is stacked in real time in the time series direction. Create a one-dimensional power data structure P_data, and store D in real time in the time series direction. s Normalized power data of the target site;

[0067] S2.2: Determine that the prediction association time is 4 hours, that is, the sliding window size w_size is 16 and the prediction time is 15 minutes. The sliding window strategy is used to construct the input-output data set for the three-dimensional matrix 3D_Data and the power data structure P_data in the time series direction. The specific method of the sliding window strategy is as follows: Take the first group of input and output data sets as an example, with 16 as the window size, the window slides in the time series direction of the three-dimensional matrix 3D_Data, and the sliding step is 1. Each time the window slides, the three-dimensional data element containing the time series is an input data. At this time, the last corresponding moment of the window is 4 hours. If the prediction target is the photovoltaic power value of site 32 in the future 15 minutes, the corresponding output data of the input data is the standardized photovoltaic power data of P_data at 4 hours and 15 minutes. The subsequent sliding windows are analogous, and each sliding generates a pair of input-output data until the corresponding moment of the output data is the last moment. Use the data set Real_Con to update the model training;

[0068] S2.3: Divide the input-output dataset into a training set and a validation set in a ratio of 10:1, namely, the training input set train_input_slices and the training output dataset train_output_slices, the validation input set vail_input_slices and the validation output dataset vail_output_slices;

[0069] S3: Integrate convolutional neural network CNN, bidirectional long short-term memory network BiLSTM and attention mechanism SE to build photovoltaic ultra-short-term power prediction model model_y;

[0070] The model includes five key modules: input data processing, abstract feature extraction, attention mechanism adjustment, time series modeling and output data processing; the input data processing module includes a sequence input layer and a sequence folding layer, which are used to receive and process data. The output of the input data processing module is connected to the feature extraction module, and the preprocessed data is sent to the convolutional neural network for feature extraction; the feature extraction module consists of a convolution layer, a batch normalization layer, an activation layer, and a pooling layer, which are used to extract spatial features. The output of the feature extraction module is connected to the attention mechanism module to adjust the model's attention to the input data and improve prediction accuracy; the attention mechanism module includes a global average pooling layer, a fully connected layer, and an activation layer. Finally, a multiplication layer is added to combine the output of the fully connected layer with the The output of the pooling layer of the convolutional neural network performs element-level multiplication operations to strengthen the focus on important features of the input data and suppress the influence of unimportant features, which is used to adjust the model's focus on the data. The output of the attention mechanism module is connected to the time series modeling module, and the adjusted feature sequence is passed to the time series modeling module for time series relationship modeling; the time series modeling module includes a sequence expansion layer, a bidirectional LSTM layer, a random drop layer, and a fully connected layer to capture time series relationships. The output of the time series modeling module is passed through the output data processing module to generate the final prediction result; the output data processing module consists of a fully connected layer and a regression layer to generate the final prediction result, ensuring that the model can fully utilize the spatiotemporal dynamic characteristics of the input data and effectively perform the prediction task;

[0071] S4: Set the photovoltaic prediction model model_y structure parameters and training hyperparameters;

[0072] The model structure parameters include 1 convolutional neural network (CNN) layer, 3x3 convolution kernel size, 32 output channels, 2 pooling size, ELU activation function parameters, 2 bidirectional long short-term memory (BiLSTM) layers, 128 and 64 hidden units, 0.2 random dropout rate, 32 channels of attention mechanism (SE), 1:8 scaling parameters, training hyperparameters include Adam optimization algorithm, 50 maximum iterations, 128 batch size, 0.001 initial learning rate, piecewise learning rate attenuation strategy, and 1 gradient threshold parameter; ensure that the model can fully utilize the spatiotemporal dynamic characteristics of the input data and effectively perform prediction tasks. The network structure diagram of the prediction model is shown in the attached figure. Figure 4 As shown;

[0073] S5: Model training: Use the structural parameters and training hyperparameters set in step S4 to train the model. The training process includes iteratively updating the parameters of the photovoltaic prediction model model_y using an optimization algorithm to minimize the loss function. The training data set is the historical input-output data set Real_Con generated in step S2.

[0074] S6: Load the three-dimensional data in the 4-hour window associated with the last moment in 3D_Data in step S2 as prediction input data, load and start the prediction model obtained in S5 to predict photovoltaic power, output the prediction value for the next 15 minutes, and denormalize it to the photovoltaic power prediction value according to the normalization parameters in step S1;

[0075] S7: Monitor the performance of the model according to the photovoltaic power prediction error evaluation index. If the monitoring error exceeds the performance threshold of 5%, return to step S5 and use the latest data set Real_Con to retrain the model. If not, continue to use the latest photovoltaic prediction model model_y to predict the output power of the target photovoltaic site.

[0076] The performance index of the photovoltaic power prediction value and the actual collected power data is calculated. The performance index uses the normalized root mean square error RMSE and the average percentage error MAPE to monitor the model performance. Taking the rated power of the photovoltaic site as the benchmark, the specific calculation of the performance index is as follows:

[0077]

[0078] Where P oi , P ei Respectively represent predicted power and actual power; P N The predicted site installed capacity is 4.9MW; N is the number of valid prediction points during the day, i.e., the number of non-zero values. The prediction effect diagram of a typical day with different weather conditions at a 15-minute step length for this specific implementation case is shown in Figure 5, and the prediction results of the model for the next 15 minutes are shown in Table 1.

[0079] Table 1 Model prediction indicators

[0080]

[0081] Example 2

[0082] The present embodiment relates to a wide-area photovoltaic site power ultra-short-term prediction device based on CNN-BiLSTM-SE, which is characterized by comprising a memory and one or more processors, wherein the memory stores executable code, and when the one or more processors execute the executable code, they are used to implement the wide-area photovoltaic site power ultra-short-term prediction method based on CNN-BiLSTM-SE of the present invention.

[0083] Example 3

[0084] The present embodiment relates to a computer-readable storage medium having a program stored thereon. When the program is executed by a processor, the method for ultra-short-term prediction of wide-area photovoltaic site power based on CNN-BiLSTM-SE of the present invention is implemented.

[0085] Those skilled in the art can understand that the above are only preferred examples of the invention and are not intended to limit the invention. Although the invention is described in detail with reference to the above examples, those skilled in the art can still modify the technical solutions recorded in the above examples or replace some of the technical features therein with equivalents. Any modification, equivalent replacement, etc. made within the spirit and principle of the invention shall be included in the protection scope of the invention.

Claims

1. A wide-area photovoltaic site power ultra-short-term prediction method based on CNN-BiLSTM-SE is characterized by: The following steps are involved: S1: Original feature data collection and preprocessing: Obtain multi-dimensional real-time data of multiple sites in the wide-area photovoltaic cluster that are independent of meteorological data, and construct an input data matrix. The data matrix of the i-th site at the k-th time is recorded as D ik , then the data matrix D ik Perform outlier and missing value processing and normalization to form a standardized matrix for the target site S2: construct dynamic three-dimensional spatiotemporal structure data 3D_Data and one-dimensional power structure data P_data, wherein the standardized data in step S1 are stacked and stored in 3D_Data in the time sequence direction. P_data is stored in stacks in the time sequence direction Normalize the power data of the target site; determine the prediction association step w_size and the prediction duration h, create input-output data with the association step w_size as the window size in the 3D_Data and P_data time series direction, and store the data set as the historical input-output data set Real_Con; S3: Integrate convolutional neural network CNN, bidirectional long short-term memory network BiLSTM and attention mechanism SE to build photovoltaic ultra-short-term power prediction model model_y, which includes five modules: input data processing, abstract feature extraction, attention mechanism adjustment, timing modeling and output data processing; the input data processing module is used to receive and process data; The abstract feature extraction module is used to extract abstract spatial features; The attention mechanism adjustment module is used to adjust the model's attention to the data. The attention mechanism module is integrated into the CNN bypass to dynamically adjust the weight of the convolutional neural network feature map to highlight key information. The time series modeling module is used to capture the time series relationship. The output data processing module is used to generate the final photovoltaic power prediction. S4: Set the structure parameters and training hyperparameters of the photovoltaic prediction model model_y, where the model structure parameters include the number of layers, convolution kernel size, number of output channels, pooling size, and activation function parameters of the convolutional neural network (CNN), the number of layers, number of hidden units, and random dropout rate parameters of the bidirectional long short-term memory network (BiLSTM), and the number of channels and scaling parameters of the attention mechanism; the training hyperparameters include the optimization algorithm, maximum number of iterations, batch size, initial learning rate, learning rate decay strategy, and gradient threshold; S5: Model training: Use the structural parameters and training hyperparameters set in step S4 to train the model. The training process includes iteratively updating the parameters of the photovoltaic prediction model model_y using an optimization algorithm to minimize the loss function. The training data set is the historical input-output data set Real_Con generated in step S2. S6: Load the three-dimensional data in the window of duration w_size associated at the last moment in 3D_Data in step S2 as prediction input data, load and start the prediction model obtained in S5 to predict photovoltaic power, output the prediction value of the future duration h, and denormalize it to the photovoltaic power prediction value according to the normalization parameter in step S1; S7: Monitor the performance of the model according to the photovoltaic power prediction error evaluation index. If the monitoring error exceeds the performance threshold σ, return to step S5 and use the latest historical data set Real_Con to retrain the model. If not, continue to use the latest photovoltaic prediction model model_y to predict the output power of the target photovoltaic site.

2. The method for ultra-short-term prediction of wide-area photovoltaic site power based on CNN-BiLSTM-SE according to claim 1 is characterized in that: The specific process of constructing the original feature data and preprocessing the data in step 1 is as follows: S2.1: Obtaining original characteristic data of several sites in the photovoltaic cluster including power data P and geographic space longitude information x and latitude information y; S2.2: Calculate the solar altitude angle h at each station at each moment. The solar altitude angle corresponding to the kth hour of each day in Beijing time is h k , the specific calculation formula is as follows: h k =sin -1 (sinδsinx+cosδcosxcos((k-12)×15)) (2) Where δ is the solar declination angle, n represents the day of the year, and x is the latitude of the station; S2.3: Integrate the data of each station at the kth time into the cluster data matrix D k , which includes the power data P of multiple photovoltaic sites, the solar altitude angle h calculated by S2.2, the longitude x and latitude y of the site, and the cluster data matrix D at the kth moment k The expression is as follows: Where D ik is the multi-dimensional feature information of the ith photovoltaic station at time k, and the power of the ith photovoltaic station at time k is P ik , the solar altitude angle is h ik , longitude and latitude are (x i ,y i ), n is the number of sites in the wide area site cluster; S2.4: For the mth target station, all stations perform difference calculations on the longitude and latitude of the target station to obtain longitude difference data Δx and latitude difference data Δy, and construct a data snapshot D' k ; Δx i =x i -x m (4) Δy i =y i -y m (5) S2.5: D' k Processing of infinite outliers and missing values ​​in the data: if the power data P of a certain station exceeds M consecutive missing values, the data of all stations in this period will be removed; if the amount of missing data is less than or equal to M, the cubic spline interpolation method is used to fill the missing values ​​to maintain the continuity and smoothness of the data; S2.6: Use the maximum value normalization method to calculate D' k Perform data normalization to obtain the standardized matrix D k * , the original data uses the difference between the maximum and minimum values ​​of each type of data as the scaling factor, and maps it to the range of (0,1) through linear transformation. The specific calculation formula is: Where x is the original data, x min is the minimum value of this type of data, x max The maximum value of this type of data, save the normalization parameter x min and x max .

3. The method for ultra-short-term prediction of wide-area photovoltaic site power based on CNN-BiLSTM-SE according to claim 1 is characterized in that: The performance monitoring of the model according to the photovoltaic power prediction error evaluation index described in step 7 is specifically carried out as follows: The performance index of the photovoltaic power prediction value and the actual collected power data is calculated. The performance index uses the normalized root mean square error RMSE and the average percentage error MAPE to monitor the model performance. Taking the rated power of the photovoltaic site as the benchmark, the specific calculation of the performance index is as follows: Where P oi , P ei Respectively represent predicted power and actual power; P N is the installed capacity of the predicted site; N is the number of valid prediction points during the day, that is, the number of non-zero values.

4. A wide-area photovoltaic site power ultra-short-term prediction device based on CNN-BiLSTM-SE, characterized in that: It includes a memory and one or more processors, wherein the memory stores executable code, and when the one or more processors execute the executable code, it is used to implement the ultra-short-term prediction method for wide-area photovoltaic site power based on CNN-BiLSTM-SE as described in any one of claims 1-3.

5. A computer-readable storage medium, characterized in that: A program is stored thereon, and when the program is executed by a processor, the ultra-short-term prediction method for wide-area photovoltaic site power based on CNN-BiLSTM-SE described in any one of claims 1 to 3 is implemented.

Citation Information

Cited By

  • Multi-station space-time correlation photovoltaic output probability prediction method and system based on interpretable deep learning

    CN119209530A