Photovoltaic power combination prediction method based on GWO-BiLSTM-Adaboost

Abstract

The invention discloses a photovoltaic power combination prediction method based on GWO-BiLSTM-Adaboost, and belongs to the technical field of new energy power generation power prediction. According to the method, the GWO algorithm, the BiLSTM neural network and the Adaboost integrated learning technology are combined, global optimization is carried out on core hyper-parameters of the BiLSTM neural network by applying the GWO algorithm, and the performance of a single prediction model is effectively improved. And then, the optimized BiLSTM neural network is taken as a basic learning unit to be incorporated into an Adaboost framework, and a unique iterative weighting strategy of the basic learning unit is utilized to finally form a strong prediction system formed by a plurality of differentiated weak classifiers. According to the method, the powerful global search capability of the GWO algorithm, the bidirectional time sequence feature extraction advantage of the BiLSTM neural network and the integrated learning characteristic of Adaboost are integrated, and the precision of photovoltaic power prediction is improved.

Description

Technical Field

[0001] The present invention relates to the technical field of renewable energy power generation prediction, and in particular to a photovoltaic power combination prediction method based on GWO-BiLSTM-Adaboost. Background Art

[0002] In recent years, solar power generation technology has garnered widespread attention in the renewable energy sector due to its advantages, including abundant resource reserves, short construction cycles, environmental friendliness, and safety and reliability. The power output of photovoltaic power generation systems exhibits significant time-varying characteristics. This volatility is primarily influenced by changes in meteorological conditions and a combination of operating parameters such as equipment temperature, voltage, and current. The volatility of photovoltaic power generation can affect the stability of the power system during grid-connected operation. To improve the safety of grid operation, it is crucial to establish a highly accurate photovoltaic power generation prediction model. From an economic perspective, accurate power forecasting can significantly reduce the electricity market risks caused by power generation fluctuations, thereby generating greater economic benefits for photovoltaic power generation companies.

[0003] Traditional LSTMs process time series data only in one direction, making them suitable for capturing historical information but unable to leverage future data, making them inadequate for predicting sudden changes in PV power. BiLSTMs, combining forward and backward LSTM layers, can simultaneously analyze historical and future time series dependencies, more comprehensively modeling the impact of meteorological factors on power, significantly improving forecast accuracy and being particularly effective in capturing complex scenarios. However, BiLSTM models still have several inherent limitations in complex tasks such as PV power forecasting. BiLSTM hyperparameters are typically tuned through manual experience or grid search, which is not only time-consuming and computationally resource-intensive but also difficult to ensure globally optimal parameter configurations and prone to falling into local minima, thus limiting model performance. GWO is a swarm intelligence optimization algorithm based on the hunting behavior of gray wolf packs. With its strong convergence and global search capabilities, combined with simple parameter setup and ease of implementation, it provides an effective approach for solving complex optimization problems. Currently, models such as LSTM and BiLSTM have demonstrated promising performance in PV power forecasting, but further improving forecast accuracy remains a core challenge in this field. Summary of the Invention

[0004] In response to the above problems, the purpose of the present invention is to overcome the limitations of traditional prediction methods and provide more reliable power prediction data support for grid scheduling and new energy management. A photovoltaic power combination prediction method based on GWO-BiLSTM-Adaboost is provided.

[0005] In order to achieve the above-mentioned purpose of the invention, the technical solution adopted by the present invention is as follows:

[0006] Step 1: Collect power data of photovoltaic power plants;

[0007] Step 2: Preprocess the data from step 1;

[0008] Step 3: Based on the data obtained in step 2, use Pearson correlation analysis to screen variables and extract features that are significantly correlated with the target variable;

[0009] Step 4: Use the GWO algorithm to solve the optimal parameters of the BiLSTM neural network model, including the learning rate, regularization coefficient, and number of hidden layer nodes;

[0010] Step 5: Use the Adaboost ensemble learning algorithm to optimize step 4 above. By integrating multiple GWO-optimized BiLSTM neural networks as weak predictors, a GWO-BiLSTM-Adaboost strong prediction model is constructed, ultimately achieving PV power generation prediction.

[0011] Step 6: Use RMSE, MAE, and MSE as evaluation indicators to evaluate the performance of the photovoltaic power prediction value obtained in step 5.

[0012] Furthermore, in step 1, a historical photovoltaic power generation data set is obtained. The data set includes 16 sets of meteorological data such as air temperature, ground pressure, relative humidity, and solar diffuse radiation index. The sampling frequency of the data set is 15 minutes, and 96 photovoltaic power observation values can be generated every day.

[0013] Furthermore, in step 2, the preprocessing of power generation data mainly includes data partitioning and data conversion. Data partitioning involves dividing the data set into a training set and a test set in a ratio of 9:1. The data conversion step involves applying extreme value normalization to the power and meteorological data, ensuring that both power and meteorological data fall within the range of 0 to 1.

[0014] The formula is:

[0015]

[0016] Where X is the original data, Y is the normalized data, min(X) and max(X) correspond to the minimum and maximum values in the data set, respectively.

[0017] Furthermore, in step 3, the Pearson correlation calculation formula is:

[0018]

[0019] Where Q is the Pearson correlation coefficient, and its value range is [-1,1]; N i 、M i are the i-th observation values of variables N and M respectively; are the sample averages of variables N and M respectively; the Pearson correlation is used to calculate the photovoltaic power data set to obtain the correlation coefficient, and the first 5 groups of data with the highest correlation are taken as input.

[0020] Furthermore, in step 4, the process of using GWO to solve the optimal parameters of the BiLSTM model includes:

[0021] (1) Initialization phase: The initial gray wolf population is constructed by random sampling, and each individual encoding represents the key hyperparameter set of BiLSTM.

[0022] (2) Fitness evaluation: For each gray wolf individual, a BiLSTM model is trained to obtain its fitness, and the top three wolves with the highest fitness (α, β, δ) are selected.

[0023] (3) Position update: Based on the position information of α, β, and δ wolves, the position coordinates of the common gray wolf are updated.

[0024] (4) Termination judgment: If the maximum number of iterations is reached, the hyperparameter configuration corresponding to α wolf (including learning rate, regularization coefficient, and number of hidden layer nodes) will be returned; if the maximum number of iterations is not reached, the iterative update will continue.

[0025] Furthermore, in step 5, the specific steps of constructing the GWO-BiLSTM-Adaboost strong prediction model are:

[0026] (1) Initialize the weight distribution of photovoltaic power training data. Initialize the network parameters and sample data, and define the photovoltaic power prediction sample as the feature vector x i and predicted label y i Randomly select N data from all samples and assign the same initial weight ω 1i =1 / N; the sample weight distribution in the first round of GWO-BiLSTM training is initialized as:

[0027] D1=(ω 11 ,ω 12 ,ω 13 ,...,ω 1n )

[0028] (2) Setting the number of weak predictors. By quantitatively analyzing how the prediction error and running time change with the number of weak predictors, the optimal number of weak predictors n for GWO-BiLSTM is determined.

[0029] (3) Weak predictor training and prediction. After the nth GWO-BiLSTM weak predictor is trained, the prediction results of the training data are output and the error sum is calculated as:

[0030] et =∑D i (i), where i = 1, 2, ... N and f(x) ≠ y

[0031] Where: f(x) is the model prediction output, and y is the true power value.

[0032] (4) Calculate the weight of the weak predictor. The weight coefficient of the current weak predictor is calculated based on the error value:

[0033] a t =1 / 2ln[(1-e t ) / e t ], where t=1,2,3,…,n

[0034] (5) Update sample weights. Adjust the training sample distribution according to the weak predictor weights:

[0035] ω t+1,i =[ω t,i exp(-a t y i f(x x ))] / Z

[0036]

[0037] Where: Z is the normalization factor.

[0038] (6) Construct a strong predictor. After completing n iterations, the weighted integration of each weak predictor is used to obtain the final strong predictor:

[0039]

[0040] Furthermore, in step 6, three key indicators are used to comprehensively quantify the photovoltaic power prediction performance, namely, the root mean square error (RMSE), the mean absolute error (MAE), and the mean square error (MSE) as evaluation indicators to evaluate the optimization ability of the photovoltaic power prediction model. The mathematical models are:

[0041]

[0042]

[0043] Where: n is the number of samples, y is the kth true value, is the k-th predicted value. The smaller the RMSE, MAE, and MSE values are, the closer the model prediction value is to the true value, and the better the prediction performance is; conversely, the larger these values are, the worse the prediction performance is.

[0044] Compared with the prior art, the technical solution of the present invention has the following beneficial effects:

[0045] This invention combines the advantages of intelligent optimization, deep learning, and ensemble learning to improve the accuracy of photovoltaic power prediction. First, GWO is used to automatically optimize BiLSTM parameters, overcoming the subjectivity of manual parameter adjustment and improving model convergence efficiency and prediction accuracy. Secondly, BiLSTM is used to parallelly learn the forward and backward time characteristics of photovoltaic data, which can comprehensively capture the temporal dynamic characteristics of key meteorological factors such as light intensity and temperature. Finally, the Adaboost ensemble learning framework is combined to adaptively weight multiple GWO-BiLSTM weak learners to enhance the model's robustness to abnormal data and make the prediction results more accurate. BRIEF DESCRIPTION OF THE DRAWINGS

[0046] The accompanying drawings are used to provide further understanding of the present invention and constitute a part of the specification. They are used to explain the present invention together with the embodiments of the present invention and do not constitute a limitation of the present invention.

[0047] Figure 1 The present invention is a flowchart of the method.

[0048] Figure 2 This is a comparison chart of photovoltaic prediction results between the present invention and the existing method.

[0049] Figure 3 The figure is a comparison chart of the RMSE evaluation indicators of the present invention and the existing method.

[0050] Figure 4 The MAE evaluation index comparison chart of the present invention and the existing method.

[0051] Figure 5 This is a comparison chart of the MSE evaluation indicators of the present invention and the existing method. DETAILED DESCRIPTION

[0052] In order to make the objectives, technical solutions and advantages of the present invention more clearly understood, the present invention is further described in detail below with reference to the accompanying drawings and embodiments.

[0053] The photovoltaic power combination prediction method based on GWO-BiLSTM-Adaboost provided by the present invention refers to Figure 1 , the method comprises the following steps:

[0054] Step 1: Collect power data of photovoltaic power plants;

[0055] Step 2: Preprocess the data from step 1;

[0056] Step 3: Based on the data obtained in step 2, use Pearson correlation analysis to screen variables and extract features that are significantly correlated with the target variable;

[0057] Step 4: Use the GWO algorithm to solve the optimal parameters of the BiLSTM neural network model, including the learning rate, regularization coefficient, and number of hidden layer nodes;

[0058] Step 5: Use the Adaboost ensemble learning algorithm to optimize step 4 above. By integrating multiple GWO-optimized BiLSTM neural networks as weak predictors, a GWO-BiLSTM-Adaboost strong prediction model is constructed, ultimately achieving PV power generation prediction.

[0059] Step 6: Use RMSE, MAE, and MSE as evaluation indicators to evaluate the performance of the photovoltaic power prediction value obtained in step 5.

[0060] In step 1, a historical photovoltaic power generation data set is obtained. The data set includes 16 sets of meteorological data, such as temperature, ground pressure, relative humidity, and solar diffuse radiation index. The sampling frequency of the data set is 15 minutes, and 96 photovoltaic power observation values can be generated every day.

[0061] In step 2, the preprocessing of power data mainly involves data partitioning and data conversion. Data partitioning involves dividing the data set into a training set and a test set at a ratio of 9:1. The data conversion step involves applying extreme value normalization to the power and meteorological data, ensuring that both data fall within the range of 0 to 1.

[0062] The formula is:

[0063]

[0064] Where X is the original data, Y is the normalized data, min(X) and max(X) correspond to the minimum and maximum values in the data set, respectively.

[0065] In step 3, the Pearson correlation calculation formula is:

[0066]

[0067] Where Q is the Pearson correlation coefficient, and its value range is [-1,1]; N i 、M i are the i-th observation values of variables N and M respectively; are the sample averages of variables N and M respectively; the Pearson correlation is used to calculate the photovoltaic power data set to obtain the correlation coefficient, and the first 5 groups of data with the highest correlation are taken as input.

[0068] In step 4, the process of using GWO to solve the optimal parameters of the BiLSTM model includes:

[0069] (1) Initialization phase: The initial gray wolf population is constructed by random sampling, and each individual encoding represents the key hyperparameter set of BiLSTM.

[0070] (2) Fitness evaluation: For each gray wolf individual, a BiLSTM model is trained to obtain its fitness, and the top three wolves with the highest fitness (α, β, δ) are selected.

[0071] (3) Position update: Based on the position information of α, β, and δ wolves, the position coordinates of the common gray wolf are updated.

[0072] (4) Termination judgment: If the maximum number of iterations is reached, the hyperparameter configuration corresponding to α wolf (including learning rate, regularization coefficient, and number of hidden layer nodes) will be returned; if the maximum number of iterations is not reached, the iterative update will continue.

[0073] In step 5, the specific steps of constructing the GWO-BiLSTM-Adaboost strong prediction model are:

[0074] (1) Initialize the weight distribution of photovoltaic power training data. Initialize the network parameters and sample data, and define the photovoltaic power prediction sample as the feature vector x i and predicted label y i Randomly select N data from all samples and assign the same initial weight ω 1i =1 / N; the sample weight distribution in the first round of GWO-BiLSTM training is initialized as:

[0075] D1=(ω 11 ,ω 12 ,ω 13 ,...,ω 1n )

[0076] (2) Setting the number of weak predictors. By quantitatively analyzing how the prediction error and running time change with the number of weak predictors, the optimal number of weak predictors n for GWO-BiLSTM is determined.

[0077] (3) Weak predictor training and prediction. After the nth GWO-BiLSTM weak predictor is trained, the prediction results of the training data are output and the error sum is calculated as:

[0078] e t =∑D i (i), where i = 1, 2, ... N and f(x) ≠ y

[0079] Where: f(x) is the model prediction output, and y is the true power value.

[0080] (4) Calculate the weight of the weak predictor. The weight coefficient of the current weak predictor is calculated based on the error value:

[0081] a t =1 / 2ln[(1-e t ) / e t ], where t=1,2,3,…,n

[0082] (5) Update sample weights. Adjust the training sample distribution according to the weak predictor weights:

[0083] ω t+1,i =[ω t,i exp(-a t y i f(x x ))] / Z

[0084]

[0085] Where: Z is the normalization factor.

[0086] (6) Construct a strong predictor. After completing n iterations, the weighted integration of each weak predictor is used to obtain the final strong predictor:

[0087]

[0088] In step 6, three key indicators are used to comprehensively quantify the photovoltaic power prediction performance, namely the root mean square error (RMSE), mean absolute error (MAE) and mean square error (MSE) as evaluation indicators to evaluate the optimization ability of the photovoltaic power prediction model. The mathematical models are:

[0089]

[0090] Where: n is the number of samples, y is the kth true value, is the k-th predicted value. The smaller the RMSE, MAE, and MSE values are, the closer the model prediction value is to the true value, and the better the prediction performance is; conversely, the larger these values are, the worse the prediction performance is.

[0091] like Figure 2-5 As shown in the figure, the GWO-BiLSTM-Adaboost method is closer to the true value under the proposed method, and the RMSE, MAE, and MSE values are all lower than those of existing prediction methods. This shows that the prediction method proposed in this paper is effective and its prediction performance is superior to that of existing LSTM, BiLSTM, and BiLSTM-Adaboost prediction methods, with higher prediction accuracy.

Claims

1. A photovoltaic power combination prediction method based on GWO-BiLSTM-Adaboost, characterized in that: The method specifically comprises the following steps: Step 1: Collect power data of photovoltaic power plants; Step 2: Preprocess the data from step 1; Step 3: Based on the data obtained in step 2, use Pearson correlation analysis to screen variables and extract features that are significantly correlated with the target variable; Step 4: Use the GWO algorithm to solve the optimal parameters of the BiLSTM neural network model, including the learning rate, regularization coefficient, and number of hidden layer nodes; Step 5: Use the Adaboost ensemble learning algorithm to optimize step 4 above. By integrating multiple GWO-optimized BiLSTM neural networks as weak predictors, a GWO-BiLSTM-Adaboost strong prediction model is constructed, ultimately achieving PV power generation prediction. Step 6: Use RMSE, MAE, and MSE as evaluation indicators to evaluate the performance of the photovoltaic power prediction value obtained in step 5.

2. A photovoltaic power combination prediction method based on GWO-BiLSTM-Adaboost according to claim 1, characterized in that: In step 1, a historical photovoltaic power generation data set is obtained. The data set includes 16 sets of meteorological data, such as temperature, ground pressure, relative humidity, and solar diffuse radiation index. The sampling frequency of the data set is 15 minutes, and 96 photovoltaic power observation values can be generated every day.

3. A photovoltaic power combination prediction method based on GWO-BiLSTM-Adaboost according to claim 1, characterized in that: In step 2, the preprocessing of power generation data mainly includes two parts: data partitioning and data conversion. The data partitioning is to divide the data set into a training set and a test set in a ratio of 9:

1. The data conversion step is to use the extreme value normalization method on the power and meteorological data so that both the power and meteorological data fall within the value range of 0 to 1. The formula is: Where X is the original data, Y is the normalized data, min(X) and max(X) correspond to the minimum and maximum values in the data set, respectively.

4. A photovoltaic power combination prediction method based on GWO-BiLSTM-Adaboost according to claim 1, characterized in that: In step 3, the Pearson correlation calculation formula is: Where Q is the Pearson correlation coefficient, and its value range is [-1,1]; N i 、M i are the i-th observation values of variables N and M respectively; are the sample averages of variables N and M respectively; the Pearson correlation is used to calculate the photovoltaic power data set to obtain the correlation coefficient, and the first 5 groups of data with the highest correlation are taken as input.

5. A photovoltaic power combination prediction method based on GWO-BiLSTM-Adaboost according to claim 1, characterized in that: In step 4, the process of using GWO to solve the optimal parameters of the BiLSTM model includes: (1) Initialization phase: construct an initial gray wolf population through random sampling, and each individual is encoded to represent the key hyperparameter set of BiLSTM; (2) Fitness evaluation: For each gray wolf individual, a BiLSTM model is trained to obtain its fitness, and the top three wolves with the highest fitness (α, β, δ) are selected. (3) Position update: Based on the position information of α, β, and δ wolves, the position coordinates of the common gray wolf are updated; (4) Termination judgment: If the maximum number of iterations is reached, the hyperparameter configuration corresponding to α wolf (including learning rate, regularization coefficient, and number of hidden layer nodes) will be returned; if the maximum number of iterations is not reached, the iterative update will continue.

6. A photovoltaic power combination prediction method based on GWO-BiLSTM-Adaboost according to claim 1, characterized in that: In step 5, the specific steps of constructing the GWO-BiLSTM-Adaboost strong prediction model are: (1) Initialize the weight distribution of photovoltaic power training data; initialize the network parameters and sample data, and define the photovoltaic power prediction sample as a feature vector x i and predicted label y i Randomly select N data from all samples and assign the same initial weight ω 1i =1 / N; the sample weight distribution in the first round of GWO-BiLSTM training is initialized as: D1=(ω 11 ,oh 12 ,oh 13 ,...,oh 1n ) (2) Setting the number of weak predictors; determining the optimal number n of weak predictors for GWO-BiLSTM by quantitatively analyzing how the prediction error and running time change with the number of weak predictors; (3) Weak predictor training and prediction: After the nth GWO-BiLSTM weak predictor is trained, the prediction results of the training data are output and the error sum is calculated as: e t =∑D i (i), where i = 1, 2, ... N and f(x) ≠ y Where: f(x) is the model prediction output, y is the true power value; (4) Calculate the weight of the weak predictor; the weight coefficient of the current weak predictor is calculated based on the error value: a t =1 / 2ln[(1-e t ) / e t ], where t=1,2,3,…,n, (5) Update sample weights; adjust the training sample distribution according to the weak predictor weights: ω t+1,i =[ω t,i exp(-a t y i f(x x ))] / Z Where: Z is the normalization factor; (6) Construct a strong predictor; after completing n iterations, the weighted integration of each weak predictor is used to obtain the final strong predictor:

7. A photovoltaic power combination prediction method based on GWO-BiLSTM-Adaboost according to claim 1, characterized in that: In step 6, three key indicators are used to comprehensively quantify the photovoltaic power prediction performance, namely the root mean square error (RMSE), mean absolute error (MAE) and mean square error (MSE) as evaluation indicators to evaluate the optimization ability of the photovoltaic power prediction model. The mathematical models are: Where: n is the number of samples, y is the kth true value, is the kth predicted value; the smaller the values of RMSE, MAE, and MSE, the closer the model predicted value is to the true value, and the better the prediction performance; conversely, the larger these values are, the worse the prediction performance.

Images