A wind and solar power generation power prediction method and related device based on LSTM-GRU-Attention

Abstract

The present invention discloses a wind-solar power prediction method and related devices based on LSTM-GRU-Attention. The prediction method comprises: obtaining historical power generation data of photovoltaic stations and wind power stations; inputting the acquired historical power generation data of photovoltaic stations and wind power stations into an optimized LSTM-GRU-Attention model, and outputting the wind-solar power generation; the LSTM-GRU-Attention model comprises a long short-term memory network module, a gated recurrent unit module and a dimensional attention module. The present invention combines a long short-term memory network LSTM with a gated recurrent unit GRU and dimensional attention, so that the LSTM-GRU-Attention method has stronger prediction ability and better robustness than several other deep learning comparison models ANN, LSTM, GRU, and LSTM-GRU.

Description

Technical Field

[0001] The present invention belongs to the technical field of wind and solar power generation prediction, and in particular relates to a wind and solar power generation power prediction method based on LSTM-GRU-Attention and related devices. Background Art

[0002] Wind and solar power generation is affected by many factors, such as weather, temperature, season, altitude, etc., which are often uncontrollable. Accurately predicting wind and solar power generation can help enterprises, governments, and users better plan and manage photovoltaic power generation systems, improve power generation efficiency and economic benefits, and also reduce the negative impact of widespread power generation uncertainty. Due to the existence of various influencing factors, the power and power generation of wind and solar power generation change dynamically over time, so accurate and stable wind and solar power generation prediction is quite difficult. Therefore, wind and solar power generation prediction has always been an important and thorny issue in the development of the entire photovoltaic and wind power industry.

[0003] In recent years, with the continuous development of deep learning and the gradual improvement of technology, more and more deep learning methods have achieved good results in various fields, and the use of deep learning methods in the field of wind and solar power generation is also increasing. However, the traditional artificial neural network ANN is a fully connected neural network that can only simulate the point-to-point mapping between the input space and the output space. In the problem of wind and solar power generation prediction, this is far from meeting the requirements because it does not consider the correlation of time series. Therefore, the recurrent neural network RNN ​​method was introduced into the prediction of wind and solar power generation. The RNN method has a network structure with recurrent connections inside, allowing the model to retain the original information. Among RNNs, the long short-term memory network LSTM is an outstanding representative. LSTM can not only handle the correlation of time series like RNN, but also prevent the problems of gradient disappearance and gradient explosion in traditional RNN. Although previous work has achieved outstanding results, it still has certain limitations for the problem of wind and solar power generation prediction. The standard LSTM method requires a large amount of data for training. Less training data will result in insufficient training of neurons, making it difficult to capture the correlation between time series, and will be affected by the mutual influence of data, resulting in inaccurate results. Summary of the invention

[0004] In order to solve the problem that wind and solar power generation cannot be accurately predicted in the prior art, the purpose of the present invention is to propose a wind and solar power generation prediction method and related devices based on LSTM-GRU-Attention, which can accurately predict photovoltaic and wind power generation.

[0005] In order to achieve the above object, the present invention adopts the following technical solutions: A wind and solar power generation power prediction method based on LSTM-GRU-Attention includes the following steps: Obtain historical power generation data of photovoltaic stations and wind power stations; The historical power generation data of photovoltaic stations and wind power stations are input into the optimized LSTM-GRU-Attention model, and the output is the predicted wind and solar power generation power; Among them, the LSTM-GRU-Attention model includes a long short-term memory network module (LSTM), a gated recurrent unit module (GRU) and a dimensional attention module (Attention).

[0006] A further improvement in the present invention is that the optimized LSTM-GRU-Attention model is obtained by the following process: Obtain historical power generation data of photovoltaic stations and wind power stations, divide them into training sets and prediction sets, and perform different time series segmentation on the data of the training sets; The segmented data is preprocessed, and the Pearson correlation coefficient is used to perform correlation analysis on the features, and the features with the highest correlation with wind and solar power generation are selected as input features; The input features are used to train the LSTM-GRU-Attention model to obtain the optimized LSTM-GRU-Attention model.

[0007] A further improvement in the present invention is that the collected data includes timestamp, temperature, humidity, wind speed, wind direction, global horizontal irradiance, array irradiance plane, wind turbine Gaia active power, solar active power, air pressure, azimuth and altitude.

[0008] A further improvement in the present invention is that the data is segmented into different time series, and the time units are divided into 1 minute, 5 minutes, 10 minutes, 15 minutes, 30 minutes and 60 minutes.

[0009] A further improvement in the present invention is that the segmented data is preprocessed, including missing value filling, outlier checking, data normalization and data standardization.

[0010] A further improvement in the present invention is that the missing value filling is specifically: using linear interpolation to fill the missing value, and the missing value is interpolated using the average value; The formula used for data normalization is:

[0011] In the formula, is the normalized data, is the original data, is the minimum value of the data, is the maximum value of the data; The formula used for data normalization is:

[0012] In the formula, is the standardized data, is the standard deviation of the data.

[0013] A further improvement in the present invention is that the input features include temperature, global horizontal radiance and array irradiance plane and wind speed.

[0014] A wind and solar power generation power prediction system based on LSTM-GRU-Attention, including: Data acquisition module, used to obtain historical power generation data of photovoltaic stations and wind power stations; The wind and solar power generation power calculation module inputs the historical power generation data of photovoltaic stations and wind power stations into the optimized LSTM-GRU-Attention model for calculation, and the output is the wind and solar power generation power; Among them, the LSTM-GRU-Attention model includes a long short-term memory network module, a gated recurrent unit module and a dimensional attention module.

[0015] An electronic device comprises a processor and a memory, wherein the processor is used to execute a computer program stored in the memory to implement the wind and solar power generation power prediction method based on LSTM-GRU-Attention.

[0016] A computer-readable storage medium, characterized in that the computer-readable storage medium has at least one instruction, and when the at least one instruction is processed and executed, the wind and solar power generation power prediction method based on LSTM-GRU-Attention is implemented.

[0017] Compared with the prior art, the present invention has the following beneficial effects: The present invention combines long short-term memory network with gated recurrent unit and dimensional attention, and utilizes the model advantages of LSTM and GRU to realize the prediction of wind and solar power generation. The method has the effectiveness and feasibility of photovoltaic and wind power generation prediction, and overcomes the shortcomings of traditional machine learning that requires a large amount of data training and lacks sufficient data, resulting in inaccurate and unstable predictions. A large number of experiments have proved that the LSTM-GRU-Attention method proposed in the present invention has stronger prediction ability and better robustness than several other deep learning comparison models ANN, LSTM, GRU, and LSTM-GRU.

[0018] Furthermore, the method compares different time series segmentations (1 minute to 60 minutes) to deduce the impact of temporal periodicity on time series data and PV power generation forecasting. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] Figure 1 This is a training flow chart of the LSTM-GRU-Attention model proposed in the present invention; Figure 2 Schematic diagram of the LSTM-GRU-Attention model proposed in the present invention; Figure 3 A comparison chart of photovoltaic power generation prediction results between the method proposed by the present invention and other methods; Figure 4 A comparison chart of wind power generation prediction results between the method proposed by the present invention and other methods; Figure 5 This is a comparison diagram of the method proposed in the present invention under different time divisions. DETAILED DESCRIPTION

[0020] In order to enable those skilled in the art to better understand the scheme of the present invention, the technical scheme in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings 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 should fall within the scope of protection of the present invention.

[0021] It should be noted that the terms "first", "second", etc. in the present invention are used to distinguish similar objects, and are not necessarily used to describe a specific order or sequence. It should be understood that the data used in this way can be interchanged where appropriate, so that the embodiments of the present invention described herein can be implemented in an order other than those illustrated or described herein. In addition, the terms "including" and "having" and any variations thereof are intended to cover non-exclusive inclusions, for example, a process, method, system, product or device that includes a series of steps or units is not necessarily limited to those steps or units that are clearly listed, but may include other steps or units that are not clearly listed or inherent to these processes, methods, products or devices.

[0022] A wind and solar power generation power prediction method based on LSTM-GRU-Attention of the present invention comprises the following steps: Step 1: Obtain historical power generation data of photovoltaic stations and wind power stations, use them as original wind and solar power generation power prediction data, and divide the data into different time series; Specifically, in step 1, 12 indicators including timestamp, temperature, humidity, wind speed, wind direction, global horizontal irradiance, array irradiance plane, wind turbine Gaia active power, solar active power, air pressure, azimuth and altitude are used to obtain the historical power generation data of photovoltaic stations and wind power stations for a total of 15 months, and form a data set, and the data set is divided into a training set and a prediction set. During the training process, the data of the first 12 months are used for training to predict the photovoltaic and wind power generation data of the last 3 months. At the same time, when the data in the training set and the prediction set are segmented into different time series, the data are segmented into 1 minute, 5 minutes, 10 minutes, 15 minutes, 30 minutes and 60 minutes according to the time unit.

[0023] Step 2: Preprocess the segmented data and use the Pearson correlation coefficient to perform correlation analysis on the features, and select the features with the highest correlation with wind and solar power generation as the input features for model training; Specifically, in the step 2, the segmented data is preprocessed, including missing value filling, outlier checking, data normalization and data standardization, etc.

[0024] The missing value filling is mainly caused by the damage of the transmission equipment. The data of the present invention is collected by sampling for 1 minute, and the time is relatively close. The measurement result may be close to the two nearby measurement values. Therefore, the present invention uses linear interpolation to fill the missing values. In order to ensure the accuracy of the data, the missing values ​​in the present invention are interpolated by the average value.

[0025] The outliers are mainly caused by weather and line damage. In the data set, an outlier should be a negative power or greater than the power of the generator. After checking, it was found that there were no outliers in the data set. Among them, the data normalization, in data normalization, the data is compressed and converted to a smaller range, and the compressed data is limited to between 0-1. This is because the difference in single data values ​​is too large, which will cause machine learning to be unable to effectively extract their features. Only after normalization can the model equally weigh all features in the data set and maintain the relevance of the data. The data normalization formula is as follows:

[0026] In the formula, is the normalized data, is the original data, is the minimum value of the data, is the maximum value of the data.

[0027] Among them, the data standardization is the process of readjusting the distribution of values ​​to ensure that the mean value of the data value is 0 and the standard deviation is 1, so that the data can be more suitable for machine learning training methods. The data standardization formula is as follows:

[0028] In the formula, is the standardized data, is the standard deviation of the data.

[0029] Specifically, in the step 2, the Pearson correlation coefficient is used to perform correlation analysis on the features, which is a statistic that measures the degree of linear correlation between two variables. Its value range is between -1 and 1. The present invention performs statistics on the important correlation of each characteristic variable in the pre-processed data for photovoltaic and wind power generation, and obtains characteristic data, including temperature, global horizontal irradiance, array irradiance plane and wind speed as the input data of the model. In the prediction of solar photovoltaic power generation, global horizontal irradiance and array irradiance plane are the most important variables, and the correlation of temperature parameters is weaker than these two. Other parameters such as wind speed, wind direction and azimuth are very low in correlation, while parameters such as humidity, air pressure and altitude are negatively correlated with photovoltaic power generation. Therefore, when training the model, the present invention only uses the data of the three characteristic parameters of temperature, global horizontal irradiance and array irradiance plane, and other data with very small correlation or negative correlation are discarded. In the prediction of wind power generation, wind speed is the most important variable, and other parameters such as temperature and wind direction are very low in correlation, while parameters such as humidity, air pressure and azimuth are negatively correlated with photovoltaic power generation. Therefore, when training the wind data prediction model, the present invention only uses the data of the characteristic parameter of wind speed as the characteristic data, and discards other data with little correlation or negative correlation.

[0030] Step 3: Initialize the LSTM-GRU-Attention model parameters such as the number of hidden neurons, learning rate, and number of training times; Specifically, in the step three, the PyTorch version used in the present invention is 1.10.1, the learning rate (LearningRate) is 1e-6, the number of hidden neurons (Hidden Layers) is 128, the convolution kernel size is 3×3, the drop_out rate is 5e-3, the batch size (Batch Size) is 128, the activation function is ReLU, the optimizer (Optimizer) is Adam, and a total of 20 epochs are trained.

[0031] Step 4: Input the obtained feature data into the LSTM-GRU-Attention model for training to obtain the optimized LSTM-GRU-Attention model; The three-month photovoltaic and wind power generation data are input into the optimized LSTM-GRU-Attention model for prediction to obtain the predicted value of wind and solar power generation.

[0032] Specifically, the specific process of the above-mentioned wind and solar power generation prediction method based on LSTM-GRU-Attention is as follows: Figure 2 As shown in Figure 1, the dataset is first divided into 1 minute, 5 minutes, 10 minutes, 15 minutes, 30 minutes and 60 minutes of data points, and significant features are extracted. The processed and segmented feature data is then fed into the LSTM-GRU-Attention model for training. The model adopts a multivariate approach and includes four input variables: the array irradiance plane (X POA )、Temperature(X T ), global horizontal radiation (X GHI ) and wind speed (X Ws ). Finally, the model evaluation indicators, namely MAE, MSE and RMSE, are used to calculate the predicted active power Y p and the real active power Y True Make a comparison.

[0033] The LSTM-GRU-Attention model is mainly composed of a long short-term memory network (LSTM) module, a gated recurrent unit (GRU) module, and a dimensional attention (Attention) module; Among them, the LSTM-GRU-Attention model training process is as follows Figure 1 As shown in the figure, specifically: the feature data is input into the LSTM module, the LSTM model extracts the sequence feature vectors with additional weights, and then these sequence feature vectors are input into the Attention module, and the relationship between the channels in these sequence feature vectors is fully mined to obtain sequence feature vectors with stronger correlation, and finally the sequence feature vectors with stronger correlation are input into the GRU module to finally obtain more accurate sequence prediction data.

[0034] Specifically, in step 4, the LSTM module includes an input gate, a forget gate, an output gate and a cell state, and the process can be described as follows: The first gate is the "Forget Gate" , which determines the information to be forgotten, and it can be expressed by the formula:

[0035] In the formula, (Sigmoid) is the activation function, is the forget gate weight parameter, is the hidden state at the previous moment, is the input data, t is the current time, is the bias parameter of the forget gate. The next step is to decide what information to add to the cell state. This step includes two steps. First, use the hidden state of the previous moment and the input data Through a gate called the input gate ( ) operation to update which information. Then use the hidden state of the previous moment and the input data Get new candidate cell information through a tanh layer , this information may be updated to the cell information. These two steps can be expressed by the formula:

[0036]

[0037] In the formula, is the input gate weight parameter, is the input gate bias parameter, is the candidate input information weight parameter, Bias parameters for candidate input information. The following will update the old cell information , becomes new cell information The update rule is to choose to forget part of the old cell information through the forget gate, and choose to add candidate cell information through the input gate. Part of the new cell information , which can be expressed as:

[0038] After updating the cell state, it is necessary to input the hidden state of the previous moment and the input data To determine which state characteristics of the output cell, the input needs to pass through a Sigmoid layer called the output gate to obtain the judgment condition, and then the cell state is passed through the tanh layer to obtain a vector with a value between -1 and 1. This vector is related to the output gate ( ) and multiply the judgment conditions to get the final output of the LSTM module. The formula can be expressed as:

[0039]

[0040] In the formula, is the output gate weight parameter, is the output gate bias parameter, is a non-linear activation function.

[0041] Specifically, the GRU module includes an update gate and a reset gate, and its process can be described as: The first door Represents the update gate, which is used to control the extent to which the state information of the previous moment is brought into the current state. The larger the value of the update gate, the more state information of the previous moment is brought in. This process is shown in the following formula:

[0042] In the formula, To update the gate weight parameters, To update the gate bias parameters.

[0043] The second door Represents the reset gate, which controls how much information of the previous state is written to the current candidate set. The smaller the reset gate, the less information of the previous state is written.

[0044]

[0045]

[0046]

[0047] In the formula, To update the gate weight parameters, To update the gate bias parameters, To update the gate weight parameters, To update the gate bias parameters, To update the gate weight parameters, To update the gate bias parameters. is a candidate hidden state, The hidden state passed to the next time step.

[0048] Specifically, the Attention module enhances the model's attention to different feature dimensions by learning the relationship and weights between features, thereby improving the model's performance on complex data. It can be expressed as:

[0049] In the formula, is the channel attention weight matrix, I is the feature tensor, To perform the maximum pooling operation on the input feature tensor I, To perform an average pooling operation on the input feature tensor I. Represents a multilayer perceptron.

[0050] Step 5: Use three loss function indicators, namely mean absolute error (MAE), mean square error (MSE) and root mean square error (RMSE), to evaluate the accuracy of the model prediction results, i.e. the predicted values ​​of wind and solar power generation. The smaller the values ​​of MAE, MSE and RMSE, the higher the accuracy of the prediction results.

[0051] Specifically, in step 5, MAE is an objective function used to measure the mean absolute error between the predicted value and the true value in the regression problem. It can measure the average error between the predicted value and the true value. The calculation formula is as follows:

[0052] In the formula, N is the total number of samples, i is the serial number of the sample, is the i-th attribute value of the true label value, is the i-th attribute value of the predicted value.

[0053] Specifically, MSE is a measure that reflects the degree of difference between the estimator and the estimated quantity, and is used to evaluate the degree of fit of the model on given data.

[0054]

[0055] Specifically, RMSE is a measurement error indicator used to measure the deviation between the observed value and the true value.

[0056]

[0057] Figure 3The results of photovoltaic power generation prediction are shown. It can be seen that as the time segmentation becomes more and more refined (60 minutes to 1 minute), the evaluation indicators MAE, MSE and RMSE (three indicators in each time period from bottom to top) all decrease to varying degrees. The LSTM-GRU-Attention method proposed in the present invention has three indicators with slightly higher values ​​than the LSTM-GRU method at 1 minute and 30 minutes. The evaluation indicators of other time periods are better than those of the ANN, LSTM, GRU, and LSTM-GRU methods. For example, at 5 minutes, the MAE, MSE, and RMSE of the LSTM-GRU-Attention method are 0.2233, 0.0842, and 0.4782, respectively; the MAE, MSE, and RMSE of the ANN method are 0.3597, 0.4253, and 0.6762, respectively; the MAE, MSE, and RMSE of the LSTM method are 0.2577, 0.2474, and 0.5635, respectively; the MAE, MSE, and RMSE of the GRU method are 0.3126, 0.2733, and 0.6200, respectively; and the MAE, MSE, and RMSE of the LSTM-GRU method are 0.2927, 0.2347, and 0.5374, respectively. By comparing the results, we can see that the LSTM-GRU-Attention method can better capture the dependencies in the time series and better solve the long-term dependency problem between data. By adding the attention mechanism, the advantages of LSTM and GRU are well combined, and the effect is better.

[0058] Figure 4 The results of wind power generation prediction are shown. It can be seen that except for four indicators with values ​​slightly higher than ANN and LSTM in 1 minute, 10 minutes and 60 minutes, the evaluation indicators in other time periods are better than the compared ANN, LSTM, GRU, and LSTM-GRU methods. For example, at 5 minutes, the MAE, MSE and RMSE of the LSTM-GRU-Attention method are 0.1693, 0.0440 and 0.3539 respectively; the MAE, MSE and RMSE of the ANN method are 0.2282, 0.1531 and 0.4948 respectively; the MAE, MSE and RMSE of the LSTM method are 0.2495, 0.2068 and 0.4924 respectively; the MAE, MSE and RMSE of the GRU method are 0.2565, 0.1498 and 0.5287 respectively; the MAE, MSE and RMSE of the LSTM-GRU method are 0.3303, 0.2013 and 0.6144 respectively. From the results, we can see that without additional parameter adjustment, the performance of the model has not decreased, which shows that the LSTM-GRU-Attention model can autonomously learn the correlation between data, which fully demonstrates that the model has good generalization ability.

[0059] Figure 5 The comparison chart shows the model performance of photovoltaic power generation prediction and wind power generation prediction in terms of evaluation indicators MAE, MSE and RMSE using the proposed LSTM-GRU-Attention model according to different time divisions. Except for the extremely high MSE value of wind power generation prediction at the 60-minute time division, which is close to 2.5, the evaluation effect of wind power generation prediction is slightly better than that of photovoltaic power generation prediction. Most of the evaluation indicators of photovoltaic power generation prediction and wind power generation prediction are below 1. Their comparison reflects the accuracy and stability of the LSTM-GRU-Attention model. All indicators at the 60-minute time division are higher than those at other time divisions. This is mainly because the data accuracy is the lowest at the 60-minute time division, and photovoltaic power generation changes rapidly over time. Changes in environmental factors within one hour will have a greater impact on photovoltaic power generation. Therefore, the effect of the 60-minute time division is the worst. Correspondingly, the performance of various measurement indicators at the 1-minute time division is better than other measurements, which also proves the conclusion that the higher the time accuracy, the better the model effect.

[0060] The present invention combines LSTM, GRU and dimensional attention to predict the accuracy of photovoltaic and wind power generation data at a certain time in the future. First, data standardization and normalization are used in the data preprocessing stage to make the model robust. Then, a correlation coefficient analysis is performed to select a suitable subset as the input of the model, which improves the availability and accuracy of the model. In addition, different time divisions are used for cross-validation training during model training: from 1 minute to 60 minutes, so that the model has long-term and short-term sequence knowledge, which enhances the robustness of the model. Finally, the effectiveness and feasibility of this method in photovoltaic and wind power generation prediction are verified by experiments: from the results, the method of the present invention has different degrees of improvement in the three evaluation indicators MAE, MSE and RMSE than the methods of ANN, LSTM, LSTM-GRU and GRU.

Claims

1. A wind and solar power generation power prediction method based on LSTM-GRU-Attention, characterized in that: The following steps are involved: Obtain historical power generation data of photovoltaic stations and wind power stations; The historical power generation data of photovoltaic stations and wind power stations are input into the optimized LSTM-GRU-Attention model, and the output is the wind and solar power generation power; Among them, the LSTM-GRU-Attention model includes a long short-term memory network module, a gated recurrent unit module and a dimensional attention module.

2. The wind and solar power generation power prediction method based on LSTM-GRU-Attention according to claim 1 is characterized in that: The optimized LSTM-GRU-Attention model is obtained through the following process: Obtain historical power generation data of photovoltaic stations and wind power stations, divide them into training sets and prediction sets, and perform different time series segmentation on the data of the training sets; The segmented data is preprocessed, and the Pearson correlation coefficient is used to perform correlation analysis on the features, and the features with the highest correlation with wind and solar power generation are selected as input features; The input features are used to train the LSTM-GRU-Attention model to obtain the optimized LSTM-GRU-Attention model.

3. The wind and solar power generation power prediction method based on LSTM-GRU-Attention according to claim 2 is characterized in that: Timestamp, temperature, humidity, wind speed, wind direction, global horizontal irradiance, array irradiance plane, wind turbine Gaia active power, solar active power, air pressure, azimuth and altitude.

4. The wind and solar power generation power prediction method based on LSTM-GRU-Attention according to claim 2 is characterized in that: The data is segmented into different time series, and the time units are divided into 1 minute, 5 minutes, 10 minutes, 15 minutes, 30 minutes and 60 minutes.

5. The wind and solar power generation power prediction method based on LSTM-GRU-Attention according to claim 2 is characterized in that: The segmented data is preprocessed, including missing value filling, outlier checking, data normalization and data standardization.

6. The wind and solar power generation power prediction method based on LSTM-GRU-Attention according to claim 5 is characterized in that: Missing value filling is as follows: missing values ​​are filled using linear interpolation, and missing values ​​are interpolated using the mean value; The formula used for data normalization is: In the formula, is the normalized data, is the original data, is the minimum value of the data, is the maximum value of the data; The formula used for data normalization is: In the formula, is the standardized data, is the standard deviation of the data.

7. The wind and solar power generation power prediction method based on LSTM-GRU-Attention according to claim 4 is characterized in that: Input features include temperature, global horizontal irradiance and array irradiance plane, and wind speed.

8. A wind and solar power generation prediction system based on LSTM-GRU-Attention, characterized in that: include: Data acquisition module, used to obtain historical power generation data of photovoltaic stations and wind power stations; The wind and solar power generation power calculation module inputs the historical power generation data of photovoltaic stations and wind power stations into the optimized LSTM-GRU-Attention model for calculation, and the output is the wind and solar power generation power; Among them, the LSTM-GRU-Attention model includes a long short-term memory network module, a gated recurrent unit module and a dimensional attention module.

9. An electronic device, characterized in that: It includes a processor and a memory, wherein the processor is used to execute a computer program stored in the memory to implement the wind and solar power generation power prediction method based on LSTM-GRU-Attention as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that: The computer-readable storage medium has at least one instruction, and when the at least one instruction is processed and executed, the wind and solar power generation power prediction method based on LSTM-GRU-Attention as described in any one of claims 1 to 7 is implemented.

Images