Photovoltaic power prediction method based on CNN-BiLSTM-Robust Attention

Abstract

The invention provides a photovoltaic power prediction method based on CNN (convolutional neural network)-BiLSTM (BiLSTM-Robust Attention), and relates to a photovoltaic power prediction method based on CNN-BiLSTM. The photovoltaic power prediction method comprises the following steps: carrying out data preprocessing on historical photovoltaic operation data and then carrying out data decomposition to obtain a plurality of intrinsic mode functions; merging the plurality of intrinsic mode functions and the historical photovoltaic operation data to obtain a reconstructed data set; using the reconstructed data set as a training sample to train a photovoltaic power prediction model, wherein the photovoltaic power prediction model comprises a CNN module, a BiLSTM module and a Robust Attention module; and inputting the real-time photovoltaic operation data into the photovoltaic power prediction model to obtain photovoltaic prediction power. The photovoltaic power prediction model can learn data space-time local features and a time series data long-term dependency relationship and enhance the model stability, and the prediction precision of the model can be improved.

Description

Technical Field

[0001] The present application belongs to the field of photovoltaic power generation technology, and specifically relates to a photovoltaic power prediction method based on CNN-BiLSTM-RobustAttention. Background Art

[0002] With the continuous expansion of the photovoltaic industry, power forecasting for large-scale photovoltaic power plants has become a hot topic of research for numerous experts and scholars both domestically and internationally. However, because photovoltaic power generation is affected by a variety of meteorological factors, such as solar radiation intensity, temperature, humidity, wind speed, and the geographical environment, its output power exhibits significant randomness and volatility, which, to a certain extent, increases the difficulty of grid dispatch. With the large-scale integration of renewable energy sources into the grid, accurate prediction of photovoltaic power has become a key technical issue in power system operation and dispatch. Currently, modeling methods based on meteorological information and historical operating data are often used to forecast photovoltaic power output. However, these methods still exhibit certain errors under complex meteorological conditions, making it difficult to meet the grid's actual demand for high-precision forecasts. Therefore, researching and implementing efficient and stable photovoltaic power forecasting methods is of great significance for optimizing power dispatch, improving grid operational safety, and enhancing resource allocation efficiency.

[0003] Existing PV power prediction methods can be broadly categorized into two main categories: physical modeling and data-driven approaches. Physical methods build theoretical simulation models based on the key design parameters of the PV system, combined with solar radiation, module performance, and meteorological conditions, to calculate the output power of the PV system. These methods offer a degree of physical interpretability and are suitable for scenarios where parameters are known and environmental conditions are controllable. However, in practice, most PV power plants are located in regions with complex geography and variable climate conditions. Their design parameters, module types, operating environments, and other factors vary significantly, making it difficult for the same physical model to maintain prediction accuracy across different scenarios, resulting in a lack of generalizability and scalability. In contrast, data-driven approaches build on historical data and incorporate statistical methods or machine learning algorithms for modeling. These methods exploit temporal features and implicit patterns in the data to predict future power. These methods do not rely on complex physical modeling processes, have low requirements for system parameters, and possess strong environmental adaptability and model transfer capabilities. Summary of the Invention

[0004] The technical problem solved by this application is: how to provide a photovoltaic power prediction method with high prediction accuracy.

[0005] This application provides a photovoltaic power prediction method based on CNN-BiLSTM-RobustAttention, which includes:

[0006] After data preprocessing, the historical photovoltaic operation data is decomposed to obtain several intrinsic mode functions;

[0007] Combining the plurality of intrinsic mode functions and the historical photovoltaic operation data to obtain a reconstructed data set;

[0008] The reconstructed data set is used as a training sample to train a photovoltaic power prediction model, wherein the photovoltaic power prediction model includes a CNN module, a BiLSTM module, and a RobustAttention module;

[0009] The real-time photovoltaic operation data is input into the photovoltaic power prediction model to obtain the photovoltaic predicted power.

[0010] Optionally, the method for preprocessing the historical photovoltaic operation data includes: cleaning and standardizing the historical photovoltaic operation data, filling in missing values ​​and detecting abnormal values.

[0011] Optionally, the data decomposition method includes:

[0012] Based on the CEEMDAN decomposition algorithm, the historical photovoltaic operation data after data preprocessing are decomposed multiple times to obtain multiple modal components and residual components.

[0013] Optionally, the method of performing multiple decompositions on the historical photovoltaic operation data after data preprocessing based on the CEEMDAN decomposition algorithm includes:

[0014] Add a white noise sequence to the historical photovoltaic operation data after data preprocessing to obtain a new signal:

[0015] w i (t)=w(t)+ε0x i (t)

[0016] Where w(t) is the historical photovoltaic operation data, ε0 is the initial white noise weight coefficient, and x i (t) is the white noise added for the i-th time, w i (t)(i=1,2,…,N) is the new signal;

[0017] For the new signal w i (t) is decomposed by EMD to obtain the first modal component IMF1(t) and the residual component r 1 (t):

[0018]

[0019] r1(t)=w(t)-IMF1(t)

[0020] Continue to add white noise to the decomposed residual component and perform EMD decomposition;

[0021]

[0022] r j (t) = r j-1 (t)-IMF j (t),(j=2,3,…,N)

[0023] Repeat the EMD decomposition until the predetermined conditions are met, and the original power data is decomposed into N modal components IMF n And the residual term R(t):

[0024]

[0025] Optionally, the method of training the photovoltaic power prediction model using the reconstructed data set as a training sample includes:

[0026] Extracting spatiotemporal local features from the reconstructed dataset using the CNN module;

[0027] Utilizing the BiLSTM module to extract long-term dependency features from the spatiotemporal local features;

[0028] Inputting the long-term dependency feature into the RobustAttention module to obtain a prediction result;

[0029] Update the network parameters of the CNN module, the BiLSTM module, and the RobustAttention module according to the prediction results;

[0030] Repeat the above steps until the training stop condition is reached.

[0031] Optionally, the method of extracting spatiotemporal local features from the reconstructed dataset using the CNN module includes:

[0032] The convolution kernel is used to perform a sliding window operation on the reconstructed data set to extract the spatiotemporal local features. The calculation formula is as follows:

[0033]

[0034] Where σ1 is the activation function, A i is the convolution layer input, w i is the weight matrix, b i is the bias matrix.

[0035] Optionally, the method of inputting the long-term dependency feature into the RobustAttention module to obtain a prediction result includes:

[0036] The initial attention weight M is obtained by multiplying the original attention score matrix A and the value matrix V (0) :

[0037]

[0038] Calculate the current weight of each position i With all input features v j The Euclidean distance D (k) ;

[0039]

[0040] In the formula, k is initially 0 and increases by 1 at each iteration;

[0041] Calculate the weight W according to the MCP function (k) , when the distance exceeds a threshold γ (default is 4.0), the weight drops to zero:

[0042]

[0043] Where, ∈ is a minimum constant;

[0044] Use softmax to normalize the weights to get The update amplitude is controlled by the step size t, the ratio of the historical value to the current weighted result is adjusted, and the aggregate representation M is updated. (k+1) :

[0045]

[0046] Repeat the iteration K times to obtain the final optimized aggregation M (K) , M (K) The result is flattened by the flatten operation and then linearly transformed to obtain the prediction result.

[0047] This application provides a photovoltaic power prediction method based on CNN-BiLSTM-RobustAttention, which has the following technical effects:

[0048] The photovoltaic power prediction model built based on the CNN module, BiLSTM module, and RobustAttention module can learn the local spatiotemporal characteristics of data, the long-term dependencies of time series data, and enhance model stability, which is conducive to improving the prediction accuracy of the model. BRIEF DESCRIPTION OF THE DRAWINGS

[0049] Figure 1 Flowchart of a photovoltaic power prediction method based on CNN-BiLSTM-RobustAttention according to one or more embodiments.

[0050] Figure 21 is an IMF graph decomposed by CEEMDAN according to one or more embodiments.

[0051] Figure 3 is a network structure diagram of a CNN module according to one or more embodiments.

[0052] Figure 4 is a structural diagram of a BiLSTM module according to one or more embodiments.

[0053] Figure 5 FIG. 4 is a schematic diagram of a mechanism of a RobustAttention module according to one or more embodiments.

[0054] Figure 6 Graphs showing prediction results of different models according to one or more embodiments. DETAILED DESCRIPTION

[0055] In order to make the purpose, technical solutions and advantages of this application more clearly understood, the following further describes this application in detail with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain this application and are not intended to limit this application.

[0056] Before describing the various embodiments of the present application in detail, the technical concept of the present application is first briefly described: the current photovoltaic power prediction method has the following problems: low prediction accuracy, weak model prediction ability, etc. To this end, the present application provides a photovoltaic power prediction method based on CNN-BiLSTM-RobustAttention. First, through data decomposition and reconstruction, the noise in the original data is removed to obtain more accurate and rich training data samples. The photovoltaic power prediction model constructed based on the CNN module, BiLSTM module, and RobustAttention module can extract and learn the local spatiotemporal characteristics of the data, the long-term dependency of the time series data, and enhance the stability of the model, which is conducive to improving the prediction accuracy of the model. The following describes the specific principles of the photovoltaic power prediction method based on CNN-BiLSTM-RobustAttention of the present application in combination with more embodiments.

[0057] Specifically, if Figure 1 The photovoltaic power prediction method based on CNN-BiLSTM-RobustAttention in the first embodiment includes the following steps:

[0058] Step S10: performing data preprocessing on historical photovoltaic operation data and then performing data decomposition to obtain a number of intrinsic mode functions;

[0059] Step S20: merging the plurality of intrinsic mode functions and the historical photovoltaic operation data to obtain a reconstructed data set;

[0060] Step S30: using the reconstructed data set as a training sample to train a photovoltaic power prediction model, wherein the photovoltaic power prediction model includes a CNN module, a BiLSTM module, and a RobustAttention module;

[0061] Step S40: input the real-time photovoltaic operation data into the photovoltaic power prediction model to obtain photovoltaic predicted power.

[0062] First, historical PV operation data is collected. For example, this data includes 10 characteristic parameters: site irradiance intensity, site temperature, angle irradiance intensity, ambient temperature, ambient humidity, wind direction, wind speed, horizontal diffuse radiation intensity, normal direct radiation intensity, and global power. The data is collected at a 5-minute interval. Next, the historical PV operation data is cleaned and standardized, with missing values ​​filled and outliers detected.

[0063] In one or more embodiments, the data decomposition method includes: performing multiple decompositions on the historical photovoltaic operation data after data preprocessing based on the CEEMDAN decomposition algorithm to obtain multiple modal components and residual components. The specific steps are as follows:

[0064] First, a white noise sequence is added to the historical photovoltaic operation data after data preprocessing to obtain a new signal:

[0065] w i (t)=w(t)+ε0x i (t)

[0066] Where w(t) is the historical photovoltaic operation data, ε0 is the initial white noise weight coefficient, and x i (t) is the white noise added for the i-th time, w i (t)(i=1,2,…,N) is the new signal;

[0067] Next, for the new signal w i (t) is decomposed by EMD to obtain the first modal component IMF1(t) and the residual component r1(t):

[0068]

[0069] r1(t)=w(t)-IMF1(t)

[0070] Then, white noise is added to the decomposed residual components and EMD decomposition is performed;

[0071]

[0072] r j (t) = r j-1 (t)-IMF j(t),(j=2,3,…,N)

[0073] Repeat the EMD decomposition until the predetermined conditions are met, and the original power data is decomposed into N modal components IMF n And the residual term R(t):

[0074]

[0075] like Figure 2 As shown in Figure 1, the original photovoltaic operation data often contains a lot of noise data, so data processing is required to improve the prediction accuracy. For example, this embodiment uses the CEEMDAN decomposition algorithm to decompose the original photovoltaic operation data into 18 IMF components (intrinsic mode functions). The results are shown in Figure 1. Figure 2 As shown in the figure, it can be seen that the IMF components are stable and no modal aliasing occurs. The IMF components are combined with the historical PV operation data set to form a reconstructed data set, which is used as a training sample. The training set and test set are divided into an 8:2 ratio.

[0076] In one or more embodiments, a method for training a photovoltaic power prediction model using a reconstructed data set as a training sample includes the following steps:

[0077] Extracting spatiotemporal local features from the reconstructed dataset using the CNN module;

[0078] Utilizing the BiLSTM module to extract long-term dependency features from the spatiotemporal local features;

[0079] Inputting the long-term dependency feature into the RobustAttention module to obtain a prediction result;

[0080] Update the network parameters of the CNN module, the BiLSTM module, and the RobustAttention module according to the prediction results;

[0081] Repeat the above steps until the training stop condition is reached.

[0082] Exemplarily, a method for extracting spatiotemporal local features from a reconstructed dataset using a CNN module includes:

[0083] The convolution kernel is used to perform a sliding window operation on the reconstructed data set to extract the spatiotemporal local features. The calculation formula is as follows:

[0084]

[0085] Where σ1 is the activation function, A i is the convolution layer input, w i is the weight matrix, b iis the bias matrix, and the network structure of the CNN module is as follows Figure 3 shown.

[0086] For example, LSTM contains three main gates: input gate, forget gate and output gate, which respectively control the memory, forgetting and output of information, and introduces cell state to store long-term information. BiLSTM is composed of two independent LSTM layers connected to each other, which process the forward and backward propagation of data respectively. Figure 4 , and the principle formulas are:

[0087] Input Gate:

[0088]

[0089] Forget Gate:

[0090] f t =σ(W f [h t-1 ,x t ]+b f )

[0091] Output gate:

[0092]

[0093] Cell status:

[0094]

[0095] Where: W is the weight matrix; h t-1 is the hidden layer state at time t-1; h t is the hidden layer state at time t; x t is the input information at time t; b is the bias term; C t-1 is the state of the memory unit at time t-1; C t is the state of the memory unit at time t.

[0096] Furthermore, the RobustAttention module is used to dynamically capture the attention relationship. Its core part is the RobustSum module, which enhances the robustness of the model to interference information by iteratively optimizing the attention weights. In the MCP (MinimaxConcave Penalty) mode, the Newton-IRLS iterative reweighted least squares method is used to update the attention weights, that is, a local upper bound is designed for the original target, and then the upper bound is optimized using the second-order Newton method. Newton-IRLS can converge quickly, usually in several iterations, such as 3 iterations. MCP is a robust loss function used to reduce the impact of outliers. In each iteration, the distance between the current estimate and the original value is calculated, and then the weight is adjusted according to the MCP function. The normalized weight is multiplied by the original attention matrix to obtain a new weight matrix. Next, the current estimate is updated by matrix multiplication. This process is iterated 3 times, and the aggregation result is gradually optimized to finally obtain the output result. The process is as follows: Figure 5 As shown in Figure 2, the processing of the RobustAttention module is as follows:

[0097] The initial attention weight M is obtained by multiplying the original attention score matrix A and the value matrix V (0) :

[0098]

[0099] Calculate the current weight of each position i With all input features v j The Euclidean distance D (k) ;

[0100]

[0101] In the formula, k is initially 0 and increases by 1 at each iteration;

[0102] Calculate the weight W according to the MCP function (k) , when the distance exceeds a threshold γ (default is 4.0), the weight drops to zero:

[0103]

[0104] Where ∈ is a very small constant, which can be set to 0.01 to prevent division by zero errors;

[0105] Use softmax to normalize the weights to get The update amplitude is controlled by the step size t, the ratio of the historical value to the current weighted result is adjusted, and the aggregate representation M is updated. (k+1) :

[0106]

[0107] Repeat the iteration K times to obtain the final optimized aggregation M (K) , M (K) The prediction result is obtained by flattening the image and then performing a linear transformation. For example, the step size t can be set to 0.1 and the number of iterations K can be set to 3.

[0108] Furthermore, after obtaining the trained photovoltaic power prediction model based on CNN-BiLSTM-RobustAttention, the test set is input into the model for prediction evaluation. The evaluation indicators include the determination coefficient R 2 , root mean square error RMSE and mean absolute error MAE, the calculation formula is:

[0109]

[0110] Where y′ i is the actual value, y i is the predicted value, is the mean of the actual values, and n is the length of the time series.

[0111] Furthermore, a comparative experiment was conducted to verify the effectiveness of the photovoltaic power prediction model based on CNN-BiLSTM-RobustAttention in improving prediction accuracy. The experiment selected BiLSTM, CNN-BiLSTM, and CNN-BiLSTM-Attention as control models, and used the data of the previous 48 hours in the test set to predict the power curve for the next 1 hour, 2 hours, 3 hours, and 4 hours respectively. The evaluation index comparison results of each model are shown in Table 1, where the bold part is the prediction result of this model, and the overall prediction effect can be seen in Figure 6 .

[0112]

[0113] Table 1 Comparison results of evaluation indicators

[0114] As shown in Table 1, the model of this embodiment shows the best results compared with other models in four different prediction time periods, and each component module such as CNN, BiLSTM, Attention and RobustAttention has improved the performance of the model to a certain extent. As can be seen from the table calculation, the MAE and RMSE of the model of this embodiment are reduced by 42.13% and 26.16% on average compared with the BiLSTM model, and R 2 The average improvement is 5.43%; compared with the CNN-BiLSTM model, the average reduction is 36.78% and 20.93%, R 2The average improvement is 3.87%; compared with the CNN-BiLSTM-Attention model, the average reduction is 15.51% and 8.27%, R 2 The average increase was 1.16%.

[0115] From the above results, we can see that this model shows better prediction accuracy under different benchmark models, and the prediction effect in different time periods is different. In order to analyze the best prediction period, the reduction of MAE, RMSE and R 2 First, from the perspective of error indicators, at 2 hours, compared with the BiLSTM model, the reductions in MAE and RMSE reached 45.35% and 28.22% respectively, which was the largest reduction; compared with the CNN-BiLSTM model, the reductions in MAE and RMSE reached 40.52% and 24.05% respectively, which was also the largest reduction; compared with the CNN-BiLSTM-Attention model, the reductions in MAE and RMSE reached 21.00% and 10.51% respectively, which was also the largest reduction. This shows that when the prediction time is 2 hours, the model has the most significant effect on suppressing errors and the best effect on improving prediction accuracy. Secondly, from the perspective of R 2 In terms of improvement, compared with BiLSTM, CNN-BiLSTM and CNN-BiLSTM-Attention models, R 2 These improvements were 5.32%, 4.05%, and 1.39%, respectively. While not the highest among all time periods, they are still significant. Combined with the significant reductions in MAE and RMSE, it can be concluded that the 2-hour model has the lowest prediction error and the best prediction stability and reliability. Therefore, this time period offers superior performance in short-term forecasting tasks and possesses high application potential.

[0116] from Figure 6 As can be seen in the figure, photovoltaic power generation shows a trend of first increasing and then decreasing over time, indicating that photovoltaic power generation is significantly affected by sunlight intensity and has a strong regularity. Comparative experiments show that when using BiLSTM alone, the model exhibits the lowest prediction accuracy in time series forecasting, indicating the limitations of the feature extraction capabilities of a single model. The introduction of the CNN-BiLSTM composite model improves prediction performance, demonstrating the complementarity between the local feature extraction capabilities of CNN and the long-term temporal dependency capabilities of BiLSTM. Incorporating the attention mechanism further improves model accuracy, demonstrating that the attention module can effectively capture the dynamic weight distribution across time steps. Using the robust attention mechanism, the model achieves optimal prediction results, verifying that the robust attention mechanism, while focusing on dynamic weights, exhibits stability under noise interference. Experimental results demonstrate that the model proposed in this embodiment demonstrates good prediction accuracy and stability.

[0117] The above describes in detail the specific implementation methods of the present application. Although some embodiments have been shown and described, those skilled in the art should understand that these embodiments can be modified and improved without departing from the principles and spirit of the present application, the scope of which is defined by the claims and their equivalents. These modifications and improvements should also be within the scope of protection of the present application.

Claims

1. A photovoltaic power prediction method based on CNN-BiLSTM-RobustAttention, characterized in that: The photovoltaic power prediction method comprises: After data preprocessing, the historical photovoltaic operation data is decomposed to obtain several intrinsic mode functions; Combining the plurality of intrinsic mode functions and the historical photovoltaic operation data to obtain a reconstructed data set; The reconstructed data set is used as a training sample to train a photovoltaic power prediction model, wherein the photovoltaic power prediction model includes a CNN module, a BiLSTM module, and a RobustAttention module; The real-time photovoltaic operation data is input into the photovoltaic power prediction model to obtain the photovoltaic predicted power.

2. The photovoltaic power prediction method based on CNN-BiLSTM-RobustAttention according to claim 1 is characterized in that: The method for data preprocessing of historical photovoltaic operation data includes: cleaning and standardizing the historical photovoltaic operation data, filling in missing values ​​and detecting abnormal values.

3. The photovoltaic power prediction method based on CNN-BiLSTM-RobustAttention according to claim 2 is characterized in that: The data decomposition method includes: Based on the CEEMDAN decomposition algorithm, the historical photovoltaic operation data after data preprocessing are decomposed multiple times to obtain multiple modal components and residual components.

4. The photovoltaic power prediction method based on CNN-BiLSTM-RobustAttention according to claim 3 is characterized in that: The method for performing multiple decompositions on the historical photovoltaic operation data after data preprocessing based on the CEEMDAN decomposition algorithm includes: Add a white noise sequence to the historical photovoltaic operation data after data preprocessing to obtain a new signal: w i (t)=w(t)+ε0x i (t) Where w(t) is the historical photovoltaic operation data, ε0 is the initial white noise weight coefficient, and x i (t) is the white noise added for the i-th time, w i (t)(i=1,2,…,N) is the new signal; For the new signal w i (t) is decomposed by EMD to obtain the first modal component IMF1(t) and the residual component r1(t): r1(t)=w(t)-IMF1(t) Continue to add white noise to the decomposed residual component and perform EMD decomposition; r j (t)=r j-1 (t)-IMF j (t),(j=2,3,…,N) Repeat the EMD decomposition until the predetermined conditions are met, and the original power data is decomposed into N modal components IMF n And the residual term R(t):

5. The photovoltaic power prediction method based on CNN-BiLSTM-RobustAttention according to claim 1, characterized in that: Methods for training photovoltaic power prediction models using reconstructed datasets as training samples include: Extracting spatiotemporal local features from the reconstructed dataset using the CNN module; Utilizing the BiLSTM module to extract long-term dependency features from the spatiotemporal local features; Inputting the long-term dependency feature into the RobustAttention module to obtain a prediction result; Update the network parameters of the CNN module, the BiLSTM module, and the RobustAttention module according to the prediction results; Repeat the above steps until the training stop condition is reached.

6. The photovoltaic power prediction method based on CNN-BiLSTM-RobustAttention according to claim 5, characterized in that: The method of extracting spatiotemporal local features from a reconstructed data set using the CNN module includes: The convolution kernel is used to perform a sliding window operation on the reconstructed data set to extract the spatiotemporal local features. The calculation formula is as follows: Where σ1 is the activation function, A i is the convolution layer input, w i is the weight matrix, b i is the bias matrix.

7. The photovoltaic power prediction method based on CNN-BiLSTM-RobustAttention according to claim 5, characterized in that: The method of inputting the long-term dependency feature into the RobustAttention module to obtain a prediction result includes: The initial attention weight M is obtained by multiplying the original attention score matrix A and the value matrix V (0) : Calculate the current weight of each position i With all input features v j The Euclidean distance D (k) ; In the formula, k is initially 0 and increases by 1 at each iteration; Calculate the weight W according to the MCP function (k) , when the distance exceeds a threshold γ (default is 4.0), the weight drops to zero: Where, ∈ is a minimum constant; Use softmax to normalize the weights to get The update amplitude is controlled by the step size t, the ratio of the historical value to the current weighted result is adjusted, and the aggregate representation M is updated. (k+1) : Repeat the iteration K times to obtain the final optimized aggregation M (K) , M (K) The result is flattened by the flatten operation and then linearly transformed to obtain the prediction result.

Images