Wind energy and solar power generation prediction model and application

Abstract

The invention discloses a wind energy and solar energy power generation prediction model and application. According to the method, a CNN (convolutional neural network) is combined with the ABILSTM to obtain a CNN-ABILSTM model, and the CNN is combined with the Transform-MLP to obtain a CNN-Transform-MLP model. The CNN can capture short-term correlation between renewable energy power generation data, and BiLSTM and Transform-MLP can utilize new information extracted by the CNN to identify inherent long-term correlation for prediction. The prediction results of the two hybrid models are more accurate and efficient, and the method has the advantages in adaptability and generalization, and has more obvious advantages in the aspect of long-term prediction.

Description

Technical Field

[0001] The invention belongs to the field of electric power systems and relates to a wind and solar power generation prediction model and application. Background Art

[0002] With the growth of electricity demand and the implementation of the "dual carbon" strategy, the power generation sector is increasingly dependent on renewable energy sources (RES). Among various types of renewable energy, wind and solar energy are particularly important due to their wide availability and contribution to sustainable energy goals. However, the intermittent and random nature of wind and solar energy makes their optimal use complicated, which in turn increases the difficulty of accurate power generation forecasting.

[0003] Forecasting of wind and solar power generation is important for reducing the deviation between power generation and demand, improving dispatch efficiency, and supporting short-term and long-term planning. RES power generation forecasts are usually divided into time intervals: ultra-short-term (within 1 hour), short-term (daily), medium-term (daily to weekly), and long-term (weekly to annual). Each forecast interval meets different operational needs - ultra-short-term forecasts are critical for the immediate management of the power grid, short-term forecasts are used for real-time dispatch and unit start and stop, medium-term forecasts help maintenance planning, and long-term forecasts are critical for strategic planning of power grid expansion.

[0004] Traditionally, independent prediction models are more favored in ultra-short-term and short-term forecasting because they can handle immediate time series patterns. However, since RES power generation is greatly affected by natural variables (such as meteorological and weather factors), it usually exhibits nonlinear and intermittent characteristics, resulting in the collected RES power generation data often containing noise and missing values ​​due to sensor failures and other reasons. Therefore, when applied to medium- and long-term forecasting, the effectiveness of independent models is subject to certain limitations.

[0005] To address these challenges, an increasing number of advanced hybrid models are being used for medium- and long-term forecasting of RES. These hybrid models combine deep learning and machine learning methods, aiming to leverage the unique strengths of each method to improve the accuracy and robustness of forecasts. Common hybrid methods include combining convolutional neural networks (CNNs), XGBoost, wavelet transforms, and multi-layer perceptrons (MLPs) with deep learning models such as bidirectional long short-term memory networks (BiLSTMs), long short-term memory networks (LSTMs), and gated recurrent units (GRUs). By combining these models, hybrid methods benefit from the complementary strengths of the two model types and address the limitations of a single approach.

[0006] Despite the proven effectiveness of hybrid models, most existing techniques focus on a single type of renewable energy (wind or solar) and are primarily optimized for short-term forecasting. While short-term forecasts are important for real-time operation of the power grid, medium- and long-term forecasts are also critical for tactical and strategic planning, as they help ensure the reliability and stability of power supply over longer periods of time. In addition, current hybrid models tend to be computationally intensive and often require extensive manual feature selection, limiting their flexibility and scalability. Summary of the invention

[0007] In view of the deficiencies of the prior art, the present invention provides a wind and solar power generation prediction model and its application.

[0008] In a first aspect, the present invention provides a wind and solar power generation prediction model, which includes a CNN-ABiLSTM hybrid sub-model and a CNN-Transformer-MLP hybrid sub-model;

[0009] The CNN-ABiLSTM hybrid sub-model includes a first CNN unit, a BiLSTM unit and an attention mechanism unit;

[0010] The first CNN unit extracts local features to obtain more effective hidden information about the renewable energy power generation data series; the BiLSTM unit further processes the local features to capture long-term and short-term dependency issues in the time series; the attention mechanism unit calculates the attention value of the feature state of data at different times to the prediction results of the renewable energy power generation data series, enhances the model's attention to important features, and obtains the optimal representation of the features.

[0011] The CNN-Transformer-MLP hybrid sub-model includes a second CNN unit, a position encoding unit and a Transformer unit;

[0012] The second CNN unit extracts local features to obtain more effective hidden information about the renewable energy generation data sequence. The output of the CNN unit passes through the position encoding unit and then enters the Transformer unit to obtain a high-dimensional feature representation of the time series data containing previous and next data information.

[0013] In a second aspect, the present invention provides the application of the above model in wind and solar power generation prediction.

[0014] The beneficial effects of the present invention are as follows: the present invention finds two hybrid models with accurate prediction results and generalization through the hybrid model method, and proposes a wind and solar power generation prediction module based on CNN-ABiLSTM and CNN-Transformer-MLP. Compared with the traditional model, the model has wider applicability and stronger generalization, and can be used for the prediction of real-time solar photovoltaic power generation data, as well as for the prediction of wind power generation data, and both short-term and long-term predictions can be performed. Specifically, the CNN-Transformer-MLP model achieves the highest accuracy in the short- to medium-term predictions of wind and solar data. On the other hand, the CNN-ABiLSTM model performs well in long-term wind power prediction, proving its superiority over other hybrid methods. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] Figure 1 It is a schematic diagram of the overall process of renewable energy power generation prediction of the present invention.

[0016] Figure 2 It is the unit diagram of a single LSTM model.

[0017] Figure 3 This is a diagram of the working principle of the BiLSTM model.

[0018] Figure 4 This is the encoder-based Transformer model diagram.

[0019] Figure 5 Construct graphs for the CNN-ABiLSTM and CNN-Transformer-MLP hybrid models.

[0020] Figure 6 This is the prediction result of the hybrid model on a one-day time scale.

[0021] Figure 7 The forecast results of the hybrid model on a time scale of one week.

[0022] Figure 8 The forecast results of the hybrid model on the time scale of January. DETAILED DESCRIPTION

[0023] The present invention will be further described below in conjunction with the accompanying drawings:

[0024] This application introduces two hybrid models: CNN-ABiLSTM and CNN-Transformer-MLP. These models are specifically designed to address the technical limitations of previous methods and provide computationally efficient, robust and versatile short-term and long-term wind and solar power generation forecasting tools. These hybrid models have been rigorously evaluated on multiple performance metrics, including root mean square error (RMSE), mean absolute error (MAE), mean square error (MSE) and coefficient of determination (R2). Experiments have shown that the proposed CNN-ABiLSTM and CNN-Transformer-MLP models consistently outperform their standalone models (CNN, BiLSTM, Transformer) in terms of accuracy and effectiveness, validating the competitiveness and robustness of these models in renewable energy generation forecasting. The results show the applicability of these models in daily, weekly, and monthly forecasts, and can become a valuable tool for immediate and strategic energy planning.

[0025] Figure 1 The overall process of renewable energy generation prediction in this application is introduced. The CNN-ABiLSTM hybrid model and the CNN-Transformer-MLP hybrid model are selected, including the following steps:

[0026] S1: Input wind and solar energy data, perform data denoising and preprocessing, divide the data into training set, validation set and test set, use the training set to train the prediction model, and use the validation set to verify after each complete traversal.

[0027] S2: After trying multiple hybrid models, we selected and established the CNN-ABiLSTM hybrid model and the CNN-Transformer-MLP model. We used the CNN neural network to detect and extract local correlations from the original time series, and then used ABiLSTM and Transformer-MLP to model the temporal dependency of the entire sequence based on these extracted features.

[0028] S3: The output processing of the BiLSTM module uses an attention mechanism model to give higher weights to more important data, while the Transformer uses a multi-head attention mechanism, and the MLP layer maps the transformed features to the final prediction.

[0029] S4: Use the CNN-ABiLSTM model and the CNN-Transformer-MLP model to obtain the renewable energy generation prediction results, and use different evaluation indicators to evaluate the model performance.

[0030] Furthermore, the sample data in step S1 is collected at intervals of 15 minutes, and the number of samples used for training and the number of samples used for testing and verification is approximately 4:1.

[0031] Data preparation is very important for building a predictive model, which involves data collection, handling missing values ​​and noise, data segmentation and normalization. Solar photovoltaic power generation and wind power generation data are greatly affected by natural factors such as weather, resulting in intermittent and volatile power generation data. The collected data often also contains missing values ​​and noise caused by sensor failures. Linear interpolation and forward or backward filling methods are used to solve missing values. The data set is then divided into three subsets for training, validating and evaluating the predictive model. The training set changes the model parameters, the validation set checks the performance, and the test set evaluates the accuracy.

[0032] Data normalization can reduce the volatility of new energy data, thereby optimizing model performance and accelerating convergence. This embodiment uses the Min-Max normalization method to scale the data in the range of 0 to 1 and is given in the following equation:

[0033]

[0034] In the formula, is the normalized value of the data, x o represents the original input, x max and x min Respectively represent the maximum and minimum values ​​of the original value.

[0035] Furthermore, the model structure constructed by S2 is as follows Figure 5 The details are as follows:

[0036] (1) CNN-ABiLSTM hybrid model

[0037] First, the preprocessed input data is processed using a CNN layer and local features are extracted to obtain more effective hidden information about the renewable energy generation data series.

[0038] Convolutional neural network (CNN) is a feedforward neural network that uses convolution operations to extract high-level features and correlations from time series. This embodiment uses two layers of 1D convolution in the CNN layer. The Con1D layer has a built-in keras.layers package, which contains 64 convolution kernels. The window size of the convolution kernel is 2, and all are activated using the Relu function. 1D convolution is a CNN structure dedicated to processing one-dimensional structured data. Its convolution kernel dimension slides in the sequence direction, which can efficiently extract local features and enhance noise resistance. This feature enables it to perform excellent results in processing complex and noisy time series tasks such as solar and wind power generation data.

[0039] The typical CNN architecture is mainly composed of: convolution layer, pooling layer, and fully connected layer. The convolution layer extracts spatial information from the data through the convolution kernel, and multiple kernels interact to generate feature maps.

[0040]

[0041] In the formula, is the jth output of the mth convolutional layer, is the i-th output of the (m-1)-th convolutional layer, Mj is the choice of input mapping, represents the weight between the i-th input and the j-th output, ⊙ represents the convolution operation, represents the j-th bias of the m layer, and the activation function (ReLU) is represented by f.

[0042] The pooling layer in this example reduces parameters by selecting the maximum value among the convolution kernel outputs, as shown below:

[0043]

[0044] The fully connected layer produces the final result of the CNN model by utilizing the output of the convolutional and pooling layers as its input, as shown below:

[0045]

[0046] In the formula, represents the final j-th output of the m-th layer of the CNN model, represents the j-th input vector of the (m-1) layer, w m Represents the weight matrix between the mth layer and the (m-1)th layer.

[0047] Secondly, the BiLSTM layer is used to further process the local features extracted by the CNN model to capture the long-term and short-term dependencies in the time series.

[0048] BiLSTM is an advanced version of the LSTM model. LSTM enhances traditional RNN by solving the gradient vanishing problem and improving long-term sequence retention through a gate protection mechanism. The LSTM unit consists of three gates: input gate, forget gate, and output gate. Its mathematical principle is as follows:

[0049]

[0050] Where U represents the weight matrix, the bias of the corresponding gate is represented by b, the symbol represents element-wise multiplication, and σ is the sigmoid activation function, which is responsible for controlling the opening and closing of the corresponding gate, which enables LSTM to capture the nonlinear characteristics of the time series. t represents the input gate, f t represents the forget gate, o t represents the output gate, c t Represents the cell state, h t Represents the hidden state. The whole process is as shown in the attached Figure 2 shown.

[0051] In the LSTM model, the forget gate uses a sigmoid function to decide whether to keep or discard the contents of the previous cell. Subsequently, the cell state is modified by merging the forget gate with the previous cell state, merging the input gate with the current input, and the output gate uses a sigmoid function to determine whether to keep or transmit the current output. The hidden state is obtained by multiplying the controlled output of the output gate by the current cell state.

[0052] The BiLSTM model consists of two LSTM modules: the forward module is used to process the forward information flow of the time series, and the reverse module is used to process the reverse information flow of the time series. The output of both modules is based on the working principle of a single LSTM model. The mathematical expression of the BiLSTM model output is as follows:

[0053]

[0054] The working operation of the BiLSTM model is as follows Figure 3 As shown in the figure, the bidirectional LSTM module can capture data information in both directions, which has significant advantages in dealing with long-term dependency issues, has certain robustness to noise and missing data, and can handle noisy data sequences such as renewable energy power generation data. Using BiLSTM to process the features and correlations of renewable energy power generation data extracted by CNN can better explore the rules hidden behind the data, thereby achieving better prediction results.

[0055] Then, the attention module is used to calculate the attention value of the feature state of the data at different times on the prediction results of the new energy power generation data sequence, so as to enhance the model's attention to important features and obtain the optimal representation of the features. The hidden layer dimension of the attention module is consistent with the dimension of the previous hidden layer.

[0056] The attention mechanism prioritizes important tasks in information overload and effectively manages limited computing resources. Its input information consists of three vectors: key (key, K), value (value, V), and query vector (query, Q). Its basic concept is mathematically expressed as follows:

[0057]

[0058] Where Atten represents the attention value, V is (K n ,V n ), (Q, K) represent the query and key vectors respectively, U represents the weight of the corresponding V, and f represents the weight transformation function.

[0059]

[0060] The self-attention mechanism benefits from three trainable parameter metrics, that is, the input sequence x is linearly mapped to V, K and Q for Softmax calculation. The multi-head attention mechanism divides a single self-attention mechanism into h parallel heads, each of which is calculated using the above formula (9) to generate a different matrix. Finally, all the head outputs of each parallel layer are concatenated and passed through a linear layer to obtain the final value. The specific process can be seen in the attached Figure 4 .

[0061] Finally, the high-dimensional calculation results are flattened through Flatten, and the results are projected to low dimensions through the fully connected Dense layer, so as to regress and obtain the prediction results of the time series data.

[0062] In summary, the CNN-ABiLSTM hybrid model is a stacked hybrid strategy that exploits the strengths of each model at different stages of prediction. The hybrid model structure starts with a 1D-CNN, followed by a BiLSTM network, and ends with an attention mechanism and a dense layer. The rolling window method is essential for structuring the input data, which allows the model to use the renewable energy data of the previous time step to predict the subsequent time step. With the help of CNN, the correlations extracted from the solar photovoltaic power generation and wind power generation data are input into the BiLSTM model, which will use the new information extracted by CNN to identify long-term correlations to predict the renewable energy generation data.

[0063] (2) CNN-Transformer-MLP hybrid model

[0064] First, the 1D convolutional layer is also used as a local feature extractor to obtain preliminary features from the input time series data.

[0065] The 1D convolution layer structure used in this model is exactly the same as that of CNN-ABiLSTM, that is, the window size of the convolution kernel is 2, and both are activated using the Relu function. Since the subsequent Transformer model also has a powerful local feature extraction capability, only one layer of 1D convolution is set here to meet the needs.

[0066] Then, the time series data is positionally encoded and input into the encoder of the Transformer model to obtain a high-dimensional feature representation of the time series data containing the previous and next data information.

[0067] The Transformer model uses a self-attention mechanism to capture the temporal characteristics and repetitive patterns in renewable energy generation data more efficiently than traditional RNNs. With position encoding, the Transformer does not need to process data sequentially like traditional RNNs, thereby improving parallel computing capabilities and learning efficiency.

[0068] The Transformer model consists of stacked encoder and decoder layers. Each layer contains a positional encoding, a multi-head attention or self-attention, and a fully connected sublayer. After layer normalization, each sublayer uses residual connections. The mathematical principle is as follows:

[0069] output sublayer =LayerNorm(x+Sublayer(x)) (10)

[0070] In the formula, output sublayer Represents the output of the sublayer, and Sublayer(x) represents the function implemented by the sublayer.

[0071] The Transformer model embeds the position information of the time series through position encoding. Specifically, the sin and cos functions are used to generate two different vectors for even and odd time steps respectively, and the position information is obtained by adding each vector and its corresponding embedding vector. The mathematical principle is as follows:

[0072]

[0073] Where P e(pos,2id) and P o(pos,2id+1) Representing even and odd position elements respectively.

[0074] After the Transformer layer in this hybrid strategy, a global average pooling layer with a single dimension is called to obtain a fixed-size output. By calculating the average value of all pixel values ​​of the feature map, each feature map is compressed into a single value, effectively reducing the number of parameters in the model. Global average pooling is more robust to spatial variations of the input by summing spatial information, and has a good effect on the prediction of solar and wind power generation data.

[0075] Global average pooling does not require parameters, avoiding overfitting in this layer. In this hybrid model, global average pooling acts as a bridge between the convolutional layer and the fully connected layer, reducing the number of layers and structural complexity of the model while maintaining or even further improving performance.

[0076] Finally, a fully connected MLP layer is used to perform nonlinear mapping on the features obtained above for feature transformation, thereby further optimizing the feature representation of the model, and the results are projected to a low dimension through a fully connected Dense layer, thereby regressing to obtain the prediction results of the acquired time series data. The dimension of this MLP is consistent with the hidden layer dimension of the Transformer.

[0077] In summary, the working principle of the CNN-Transformer-MLP hybrid model is similar to that of the CNN-ABiLSTM. After extracting the high-level features and correlations of the data with CNN, these feature information is processed by the position encoding layer, and the output of the Transformer model is then fed into the MLP to produce the final result. The key difference between the two hybrid models is the use of BiLSTM and Transformer models. The BiLSTM model is susceptible to the gradient vanishing problem, which may hinder its ability to model long-range dependencies present in RES data. Since renewable energy generation data is sequential in nature, the BiLSTM model requires additional training time. The encoder-based Transformer model uses its powerful parallel processing capabilities to deal with the above problems. This method takes advantage of the different advantages of different models by mixing models, taking advantage of each.

[0078] Furthermore, the main function of the attention mechanism in S3 is to assign more weights to more meaningful time steps in the output layer of BiLSTM, enhance the model's attention to key information, dynamically adjust the weight of the input content to help ignore some irrelevant information, enable the model to effectively adapt to the variable length of the mobile mode and time dependency, and ensure that the BiLSTM layer can selectively focus on more relevant parts of the time series during the prediction process, which has a significant effect when analyzing the hidden laws behind complex solar photovoltaic power generation data and wind power generation data. The CNN-ABiLSTM hybrid model adopts a simple attention mechanism and abandons the self-attention or multi-head attention mechanism. The mathematical principle of this simple attention mechanism is as follows:

[0079]

[0080] The new set of values ​​denoted by e is created by computing the dot product of the input sequence x and the weight matrix W, and then passing it through the activation function tanh. β is the probability distribution of e, and context is the i-th input sequence (x i ) and the i-th probability distribution (β i ) is the weighted sum of .

[0081] The prediction results of the two hybrid models are better than those of the single model, and the two hybrid models have their own advantages and disadvantages, which depend on many factors such as the application environment. The use of hybrid models can take advantage of the strengths of each model and better solve the prediction problem of renewable energy power generation data.

[0082] Furthermore, four separate evaluation indicators are used in S4 to evaluate the prediction model. The evaluation indicators are measured by calculating the difference between the true value and the predicted value. The smaller the value of the evaluation indicator, the higher the accuracy of the method: mean absolute error (MAE), root mean square error (RMSE), mean square error (MSE) and final determination coefficient (R 2 ), the mathematical formula is as follows:

[0083]

[0084] Where N represents the total number of data points, Represents the predicted value of renewable energy power generation data, y R is the original data, R ss represents the residual sum of squares, T ss is the sum of the squares.

[0085] In summary, this application studies a series of hybrid framework strategies for renewable energy generation forecasting, including CNN-LSTM, CNN-BiLSTM, CNN-ABiLSTM, and CNN-Transformer-MLP models. Among them, CNN-ABiLSTM and CNN-Transformer-MLP show superior performance due to their better handling of the nonlinear and intermittent characteristics of renewable energy generation data, and always outperform other hybrid techniques. Specifically, the CNN-Transformer-MLP model achieved the highest accuracy in short- to medium-term forecasts of wind and solar data. On the other hand, the CNN-ABiLSTM model performed well in long-term wind energy forecasting, demonstrating its advantages over other hybrid methods.

[0086] Application Examples

[0087] In order to verify the effectiveness of the wind and solar energy prediction method based on the hybrid model described in this application, the renewable energy data from the ENTSO-E platform with an interval of 15 minutes covering the Germany and Luxembourg regions from 2018 to 2021 was used as the training set, and the renewable energy data from 2021 to 2023 was used for verification.

[0088] In the CNN-ABiLSTM hybrid model, the CNN consists of two Con1D layers, using the built-in keras.layers package with 64 convolution kernels, a kernel size of 2, and ReLU as the activation function. The BiLSTM layer consists of 50 LSTM units, each of which is activated by ReLU. Table 1 summarizes the configuration of different layers.

[0089] In the CNN-Transformer-MLP hybrid model, the CNN layer includes a Conv1D layer with 16 convolution kernels, each kernel size is 1, and ReLU is used as the activation function. The optimal hyperparameters of the hybrid model obtained by testing are shown in Table 2.

[0090] Table 1. Hyperparameters of the CNN-ABiLSTM model

[0091]

[0092]

[0093] Table 2. Hyperparameters of the CNN-Transformer-MLP model

[0094]

[0095] Figure 6 , Figure 7 and Figure 8 The following are the prediction results of the hybrid model on the time scales of one day, one week, and one month, respectively. As can be seen from the figure, for the daily prediction results of solar energy data, the differences among all models are not large and are all within an acceptable range; while for the prediction results of wind energy data, the prediction results of the CNN model are poor. The main factor leading to its poor performance is that it cannot effectively handle the inherent complex patterns behind wind data. For the power generation forecast for the next week, the CNN-Transformer-MLP hybrid model shows relatively good performance for both data sets, while the CNN-ABiLSTM model outperforms other models in the prediction results for the next month. Tables 3-5 describe in detail the evaluation indicators of the prediction performance of the hybrid model and the baseline model in three different periods.

[0096] Table 3. Comparison of daily forecast results of different models

[0097]

[0098] Table 4. Comparison of weekly forecast results of different models

[0099]

[0100] Table 5. Comparison of monthly forecast results of different models

[0101]

[0102] The data show that both hybrid models outperform the baseline model in the daily forecast results of renewable energy data, with RMSE values ​​of 36.94 and 31.32 for solar data prediction and 44.82 and 43.07 for wind data prediction.

[0103] Specifically, the CNN-Transformer-MLP model significantly outperforms other models in daily forecasting, the CNN-Transformer-MLP model still slightly outperforms CNN-ABiLSTM in weekly forecasting of solar power generation, and the CNN-Transformer-MLP model outperforms other models in weekly forecasting of wind power generation. For monthly forecasting, the CNN-ABiLSTM model performs better than other models.

[0104] In particular, it can be seen from Table 3 that the CNN-Transformer-MLP model outperforms the CNN-ABiLSTM model, with lower MSE, MAE, and RMSE values. Compared with other hybrid and baseline models, the results of the CNN-Transformer-MLP model in the prediction of the next week show that its performance and efficiency are in a leading position in all aspects. In the prediction of the next month, the CNN-ABiLSTM model performs slightly better overall. The CNN-Transformer-MLP model is more accurate in predicting the solar data set, with a RMSE value of 54.89 lower than that of the CNN-ABiLSTM. During the prediction of the wind power data set, the CNN-ABiLSTM model performs better. This situation emphasizes the importance of using the CNN-ABiLSTM model for long-term wind power forecasting.

[0105] It can be seen from the above application examples that the wind and solar power generation prediction method based on CNN-ABiLSTM and CNN-Transformer-MLP described in this application can simultaneously meet the needs of solar energy data and wind energy data prediction, and can take into account long-term predictions at the same time. It can be applied to the prediction of renewable energy power generation and promoted to new power systems at all levels with large-scale access to new energy, to promote the smooth operation of the power system and improve the efficiency of power dispatching.

[0106] The above description is only a preferred embodiment of the present application and is not intended to limit the present application. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present application shall be included in the scope of protection of the present application.

Claims

1. A wind and solar power generation prediction model, characterized in that: The model includes a CNN-ABiLSTM hybrid sub-model and a CNN-Transformer-MLP hybrid sub-model; The CNN-ABiLSTM hybrid sub-model includes a first CNN unit, a BiLSTM unit and an attention mechanism unit; The first CNN unit extracts local features to obtain more effective hidden information about the renewable energy generation data sequence; The BiLSTM unit further processes the local features to capture the long-term and short-term dependency problems in the time series; the attention mechanism unit calculates the attention value of the feature states of data at different times to the prediction results of the new energy power generation data series, enhances the model's attention to important features, and thus obtains the optimal representation of the features; The CNN-Transformer-MLP hybrid sub-model includes a second CNN unit, a position encoding unit and a Transformer unit; The second CNN unit extracts local features to obtain more effective hidden information about the renewable energy generation data sequence. The output of the CNN unit passes through the position encoding unit and then enters the Transformer unit to obtain a high-dimensional feature representation of the time series data containing previous and next data information.

2. A wind and solar power generation prediction model according to claim 1, characterized in that: The first CNN unit uses two layers of 1D convolution, the window size of the convolution kernel is 2, and both are activated using the Relu function.

3. A wind and solar power generation prediction model according to claim 1 or 2, characterized in that: The CNN-ABiLSTM hybrid sub-model also includes a Flatten layer and a Dense layer. The high-dimensional calculation results are flattened by the Flatten layer, and the results are projected to low dimensions by the fully connected Dense layer, so as to regress and obtain the prediction results of the time series data.

4. A wind and solar power generation prediction model according to claim 1, characterized in that: The second CNN unit uses a single-layer 1D convolution, the window size of the convolution kernel is 2, and the Relu function is used for activation.

5. A wind and solar power generation prediction model according to claim 1 or 4, characterized in that: The CNN-Transformer-MLP hybrid sub-model also includes an MLP layer and a Dense layer. A fully connected MLP layer is used to perform nonlinear mapping on the acquired features for feature conversion, thereby further optimizing the feature representation of the model, and the results are projected to a low dimension through a fully connected Dense layer, thereby regressing to obtain the prediction results of the acquired time series data.

6. A wind and solar power generation prediction model according to claim 1, characterized in that: It also includes four separate evaluation metrics to evaluate the prediction model: mean absolute error, root mean square error, mean square error, and final determination coefficient.

7. A wind and solar power generation prediction model according to claim 1, characterized in that: It also includes inputting wind and solar energy data, performing data denoising and preprocessing, dividing the data into training, validation and test sets, using the training set to train the prediction model, and using the validation set to verify after each complete traversal.

8. A wind and solar power generation prediction model according to claim 7, characterized in that: The data preprocessing includes using linear interpolation and forward or backward filling to solve missing values, and also includes normalizing the data.

9. Application of the prediction model according to any one of claims 1 to 8 in wind and solar power generation prediction.

Images