CEEMDAN-DBO-BiLSTM-based photovoltaic output interval prediction method

Abstract

The invention provides a photovoltaic output interval prediction method based on CEEMDAN-DBO-BiLSTM, and the method comprises the following steps: S1, decomposing a photovoltaic power sequence into a plurality of mode components through an adaptive noise complete set empirical mode decomposition method according to the nonlinear and non-stationary characteristics of distributed photovoltaic historical data; s2, carrying out secondary decomposition on the high-frequency non-stationary component obtained by primary decomposition, and reconstructing all components into a trend component and an oscillation component by adopting a sample entropy method; s3, optimizing parameters of the bidirectional long-short-term memory network based on a dung beetle optimization algorithm, and finally obtaining distributed photovoltaic point predicted values of the two components; and S4, probability density estimation is performed on the point prediction error of the oscillation component based on a kernel density estimation method, and the point prediction values are superposed to obtain an overall prediction interval result, so that the complexity of the original prediction component is significantly reduced, and the prediction precision of the model is further improved.

Description

Technical Field

[0001] The present invention relates to the field of smart grid prediction technology, and in particular to a photovoltaic output interval prediction method based on CEEMDAN-DBO-BiLSTM. Background Art

[0002] With the continuous advancement of distributed photovoltaic power generation technology and the expansion of PV power plants, distributed photovoltaic power generation output forecasting has become a key component in ensuring efficient operation and optimized scheduling of PV power plants. It has also become a widely used technology in smart grids, with significant practical significance for improving power generation efficiency and reducing operating costs. Accurately predicting changes in renewable energy generation allows for timely adjustments to scheduling plans, thereby reducing system reserves and the resulting operating costs. However, due to various factors such as meteorological conditions and environmental fluctuations, the output of distributed photovoltaic power generation often exhibits significant volatility and uncertainty. This characteristic poses challenges to the stable operation and accurate forecasting of power generation systems.

[0003] Therefore, a scientific and rational distributed photovoltaic power generation output forecasting method not only provides a solid theoretical basis for the adaptive planning and construction of photovoltaic systems, but also plays an important role in optimizing the operation of distribution networks. Through accurate forecasting and analysis, the economic benefits of photovoltaic power generation can be significantly improved while ensuring grid stability and efficient operation, thereby promoting the comprehensive utilization and sustainable development of clean energy.

[0004] For example, patent application number 202311635421.7, published on October 22, 2024, discloses a photovoltaic power prediction method based on similarity reorganization. The method first obtains time series of photovoltaic power, total irradiance, diffuse irradiance, temperature, and humidity. These time series are then decomposed into multiple intrinsic mode functions (IMFs) using CEEMDAN. The DTW algorithm is used to calculate the similarity between the IMFs, and those with high similarity are reorganized. The Adam optimizer is used to optimize the network parameters of the LSTM model. The improved DBO optimization algorithm is then used to optimize the hyperparameters of the LSTM model, resulting in an optimized LSTM prediction model. The reorganized IMF data is normalized and then fed into the LSTM optimized prediction model to obtain prediction results for each reorganized sequence. These prediction results are then reorganized to achieve the final prediction. This method reduces the non-stationarity of the photovoltaic power time series, enhancing prediction accuracy and reducing computational complexity.

[0005] While the improved DBO optimization algorithm described in the aforementioned literature can theoretically optimize LSTM hyperparameters, since it optimizes a unidirectional LSTM network, its optimization process depends on the quality of the initial population, the number of iterations, and the parameters of the chaotic map. Improper settings of these parameters can cause the optimization process to become stuck in a local optimum, preventing the global optimal hyperparameters from being found. This can lead to inaccurate predicted values. Furthermore, photovoltaic power generation is strongly affected by weather conditions. The components decomposed solely through the CEEMDAND method often contain significant high-frequency nonstationary noise, creating instability. Recombining data using similarity complicates the entire calculation process. If a regional value contains errors, directly removing the vibration value can lead to a higher number of vibration values ​​in severe weather conditions. Directly removing these values ​​may prevent the CEEMDAND decomposition and DTW recombination from reducing the nonstationarity of the time series, resulting in decreased prediction accuracy. Summary of the Invention

[0006] The present invention provides a photovoltaic output interval prediction method based on CEEMDAN-DBO-BiLSTM, which not only significantly reduces the complexity of the original prediction component in the smart grid, but also further improves the prediction accuracy of the model.

[0007] To achieve the above objectives, the technical solution of the present invention is: a photovoltaic output interval prediction method based on CEEMDAN-DBO-BiLSTM, comprising the following steps: S1 acquiring photovoltaic data; and decomposing the photovoltaic power series into two or more modal components using an adaptive noise complete set empirical mode decomposition method based on the nonlinear and non-stationary characteristics of the photovoltaic historical data, wherein the modal components include high-frequency non-stationary components; S2 uses the sample entropy method to perform secondary modal decomposition on the high-frequency non-stationary components obtained by the primary decomposition, and reconstructs the secondary components into trend components and oscillation components; the sample entropy method quantifies the complexity of each high-frequency non-stationary component and selects the disturbance component with the largest entropy value for secondary modal decomposition; S3 optimizes the parameters of the bidirectional long short-term memory network based on the dung beetle optimization algorithm, and finally obtains the distributed photovoltaic point prediction values ​​of the two components; S4 performs probability density estimation on the point prediction error of the oscillation component based on the kernel density estimation method to obtain an estimated value, and superimposes the distributed photovoltaic point prediction value of the two components with the estimated value to obtain the overall prediction interval result.

[0008] In the above setup, the photovoltaic power series is first decomposed into multiple modal components using the complete ensemble empirical mode decomposition with adaptive noise (CEEMDAN) method. Secondly, the high-frequency non-stationary components obtained from the primary decomposition are subjected to a secondary decomposition to obtain secondary components. The sample entropy method is then used to reconstruct all components into trend components and oscillation components to quantify the complexity of each component. This indicator is positively correlated with the randomness of the sequence, and the larger the entropy value, the more irregular the signal. The perturbation component with the largest entropy value is then selected for secondary modal decomposition, parsing it into reconstructed components with clear periodic characteristics and enhanced stationarity. This can eliminate the problem of inaccurate components caused by significant high-frequency non-stationary noise introduced by the high-order intrinsic mode function generated by DEEMDAN. The secondary modal decomposition is performed only on the perturbation component with the largest entropy value. The calculation method is simple and does not require a large number of similarity algorithms to calculate and determine. The bidirectional long short-term memory (BiLSTM) network is then optimized using the Dung Beetle Optimization (DBO) algorithm. The parameters are optimized, and finally the distributed photovoltaic point prediction values ​​of the two components are obtained; finally, the probability density estimation of the point prediction error of the oscillation component is performed through the kernel density estimation method (KDE), and the point prediction values ​​are superimposed to obtain the overall prediction interval result. The bidirectional long short-term memory network can associate the output value with the positive and reverse conditions of the input value, making the optimal parameters more accurate. At the same time, after obtaining the distributed photovoltaic point prediction value, it is necessary to determine the probability density estimation of the point prediction error through the kernel density estimation method to obtain the estimated value. The final estimated value is limited by the area with the largest probability, and the prediction interval is narrowed. At the same time, the accuracy of the entire interval can also be made higher, which can provide new ideas for the prediction of distributed photovoltaic output interval. The obtained prediction results have important engineering practical value for ensuring the safe and stable operation of the power grid.

[0009] Furthermore, the steps of decomposing the photovoltaic power sequence into two or more modal components using the adaptive noise complete set empirical mode decomposition method include: S11: Add to The Gaussian white noise with sub-mean of 0 is constructed The sequence to be decomposed in the experiment Is the signal sequence number: ,in is the preset Gaussian white noise weight coefficient; For the Gaussian white noise generated during the secondary processing; S12 pair sequence Perform EMD decomposition to obtain the first modal component IMF1 and take its mean as the first IMF obtained by CEEMDAN decomposition. The first modal component obtained by decomposition; Represents the residual signal after the first decomposition.

[0010] With the above settings, CEEMDAN can effectively reduce modal aliasing and improve the stability and reliability of the decomposition by adding Gaussian white noise to the signal and performing multiple EMD decompositions, and then taking the mean of the modal components.

[0011] Furthermore, after step S12, the following steps are further included: The first After adding specific noise to the residual signal, continue to perform EMD decomposition. , Quantity; Denotes the adaptive noise complete set empirical mode decomposition method for the The weight coefficient of the noise added to the residual signal of the stage; S14 If the stopping condition of the adaptive noise complete set empirical mode decomposition method is met, The residual signal of the sub-decomposition If it is a monotonic signal, the iteration stops and the decomposition of the adaptive noise complete set empirical mode decomposition method algorithm ends.

[0012] With the above settings, the complete ensemble empirical mode decomposition (CEEMDAN) with adaptive noise adds noise to perform EMD decomposition after the jth stage decomposition until the jth stage. The residual signal of the sub-decomposition The decomposition process stops only when the signal becomes monotonic, thus ensuring that the data maintains the same robustness after multiple decompositions, effectively reducing the signal reconstruction error and significantly improving the decomposition efficiency.

[0013] Furthermore, the step of reconstructing all components into trend components and oscillation components using the sample entropy method includes: S21 optimizes the bandwidth constraint by variational mode decomposition (VMD), which can decompose high-frequency non-stationary components into more stable modal components. The VMD constrained variational model constructed in step S21 is as follows: ; are the submodal component set and its corresponding center frequency set respectively; is the pre-set or estimated number of modes; is the partial derivative operator; * is the convolution operator; is the unit pulse function; represents an imaginary number; is the signal to be decomposed; S22 adds penalty items , Lagrange multiplication operator , transformed into an unconstrained variational problem: ; The alternating direction multiplication method is used to Optimize and output the decomposed components when the required accuracy is achieved; S23 performs a sample entropy operation on the decomposed components that have reached the required accuracy output.

[0014] With the above settings, analysis of the CEEMDAN decomposition results shows that the high-order intrinsic mode functions (IMFs) produced often contain significant high-frequency non-stationary noise. Therefore, an adaptive secondary decomposition strategy is adopted. First, the complexity of each component is quantified using the SE method. This metric is positively correlated with sequence randomness, with larger entropy values ​​indicating more irregular signals. Secondary modal decomposition is then performed on the perturbation component with the largest entropy value, parsing it into reconstructed components with clear periodic characteristics and enhanced stationarity. This hierarchical processing mechanism effectively addresses the error accumulation effect caused by high-frequency direct prediction and the challenges posed by non-stationary signals to the adaptability of the prediction model, thereby systematically improving the accuracy of time series prediction.

[0015] Furthermore, step S3 includes: S31 LSTM mainly controls data transmission through three gates, namely, forget gate, input gate, and output gate; S32 uses a bidirectional long short-term memory network composed of two forward and reverse LSTM networks. It can make bidirectional predictions based on the changing patterns of data and use the bidirectional long short-term memory network to predict photovoltaic power. The S33DBO algorithm constructs a multimodal interaction model by analyzing the typical behaviors of dung beetle populations, such as ball navigation, courtship dance, food optimization, competitive stealing, and reproductive migration. S34 and DBO algorithms are used to optimize the hyperparameters of bidirectional long short-term memory networks.

[0016] In the above settings, the input value is determined bidirectionally through the bidirectional long short-term memory network, and then the hyperparameter optimization is performed through the DBO algorithm.

[0017] Furthermore, step S34 includes: (1) Initialize various parameters in the DBO optimization algorithm, including the number of iterations, dung beetle size, upper and lower limits of the optimization variable size, and the maximum number of iterations; (2) Input the data set, divide it into training set and test set, calculate the individual fitness of the DBO algorithm, and save the optimal individual and the optimal position; (3) If the optimal solution can be determined when the termination conditions are met, the program exits; otherwise, it returns to step (2); (4) The optimized parameters are reassigned to the bidirectional long short-term memory network, and the network is trained to obtain the optimized photovoltaic point prediction results.

[0018] With the above settings, the parameters of the LSTM network are optimized based on the DBO algorithm, and finally the distributed photovoltaic point prediction values ​​of the two components are obtained.

[0019] Furthermore, the kernel density estimation method in step S4 includes: S41 determining the kernel density; ,in, is bandwidth; is the total number of quantiles; is a dataset consisting of conditional quantiles; is the kernel function; The bandwidth is The kernel density expression when ; S42. The formula of the kernel function is as follows: ; Where x is the sample data, h is the bandwidth parameter of the kernel function; S43, superposition prediction interval: The error probability density function of the oscillation component is calculated by the kernel density estimation method, and its integral is obtained Probability distribution function; at the significance level of If the error sample falls into the error interval The probability is not less than , then the interval It is called confidence The confidence interval of the oscillation component is superimposed with the point prediction value of the trend component and the oscillation component to obtain the final prediction interval. The calculation formula is as follows: ; in, is the upper limit of the prediction interval, is the lower limit of the prediction interval; is the point forecast value of the superposition of trend component and oscillation component, is the rated power of distributed photovoltaic.

[0020] In the above settings, the KDE method first determines the kernel density function and the kernel density formula, thereby making the probability distribution function of the kernel density function, and then determines the probability of the confidence interval, and then determines the prediction interval. Its application in the prediction interval calculation can provide more accurate and reliable prediction results. BRIEF DESCRIPTION OF THE DRAWINGS

[0021] Figure 1 The figure is a flow chart of the distributed photovoltaic output interval prediction method based on CEEMDAN-DBO-BiLSTM of the present invention.

[0022] Figure 2 This is a diagram of the BiLSTM working principle of the present invention.

[0023] Figure 3 This is a flowchart of the present invention's optimization of the BiLSTM model based on the DBO algorithm. DETAILED DESCRIPTION

[0024] Example 1.

[0025] like Figure 1-3 As shown in FIG, a distributed photovoltaic output interval prediction method based on CEEMDAN-DBO-BiLSTM includes the following steps: Based on the nonlinear and non-stationary characteristics of distributed photovoltaic historical data, S1 decomposes the photovoltaic power series into multiple modal components using the Complete Ensemble Empirical Mode Decomposition with Adaptive Noise (CEEMDAN) method, thereby effectively extracting useful information from the signal; S2 performs a secondary decomposition on the high-frequency non-stationary components obtained from the primary decomposition, and uses the sample entropy (SE) method to reconstruct all components into trend components and oscillation components; S3 optimizes the parameters of the bidirectional long short-term memory network (BiLSTM) based on the Dung Beetle Optimization Algorithm (DBO), and ultimately obtains the distributed photovoltaic point prediction values ​​of the two components; S4 performs probability density estimation on the point prediction error of the oscillation component based on the kernel density estimation method (KDE), and superimposes the point prediction values ​​to obtain the overall prediction interval results.

[0026] The steps for calculating the complete ensemble empirical mode decomposition of adaptive noise (CEEMDAN) are as follows: S11 transforms the signal to be decomposed into Add to The Gaussian white noise with sub-mean of 0 is constructed The sequence to be decomposed in the experiment ,in is the preset Gaussian white noise weight coefficient; Gaussian white noise generated during the secondary processing; S12 for the above sequence Perform EMD decomposition to obtain the first modal component (IMF1) and take its mean as the first IMF obtained by CEEMDAN decomposition. ,in, represents the first modal component obtained by CEEMDAN decomposition; represents the residual signal after the first decomposition. By adding Gaussian white noise to the signal and performing multiple EMD decompositions, and then taking the mean of the modal components, CEEMDAN can effectively reduce modal aliasing and improve the stability and reliability of the decomposition.

[0027] After adding specific noise to the j-th stage residual signal obtained after S13 decomposition, EMD decomposition is continued. ; ; ; Indicates that CEEMDAN preset The weight coefficient of the noise added to the residual signal of the stage; S14 If the EMD stop condition is met, the residual signal of the nth decomposition If the signal is monotonic, the iteration stops and the CEEMDAN algorithm decomposition ends. While maintaining the same robustness, the CEEMDAN algorithm effectively reduces the signal reconstruction error and significantly improves the decomposition efficiency.

[0028] The specific steps of using the sample entropy (SE) method in step S2 to reconstruct all components into trend components and oscillation components are as follows: S21 optimizes the bandwidth constraint through variational mode decomposition (VMD), which can decompose high-frequency non-stationary components into more stable modal components. The VMD constrained variational model constructed in step S21 is as follows: ; in, are the submodal component set and its corresponding center frequency set respectively; is the pre-set or estimated number of modes; is the partial derivative operator; is the convolution operator; is the unit pulse function; Represents a preset imaginary number; is the signal to be decomposed; is a constraint condition.

[0029] S22 adds penalty items , Lagrange multiplication operator , transformed into an unconstrained variational problem: ; L is the Lagrangian function; the alternating direction multiplication method is used to Optimize and output the decomposed components when the required accuracy is achieved; S23 performs a sample entropy operation on the decomposed components that have reached the required accuracy output, where the sample entropy calculation formula is as follows: ; in ,Right now ; in is the embedding dimension, usually 1 or 2.

[0030] (2) ,Right now ; (3) Setting similarity tolerance ,but The number of ; in, The similarity tolerance is usually 0.1 to 0.25 times the variance of the data to be calculated, and the reference value is 0.25 times. For two sequences in Next match The probability of a sample point is ; (4) Extend the vector dimension to , the sample entropy is ; in, Probability; Analysis of the CEEMDAN decomposition results shows that the high-order intrinsic mode functions (IMFs) it produces often contain significant high-frequency non-stationary noise. Therefore, an adaptive secondary decomposition strategy is adopted: first, the complexity of each component is quantified using the SE method. This indicator is positively correlated with the randomness of the sequence (the larger the entropy value, the more irregular the signal). Then, the disturbance component with the largest entropy value is selected for secondary modal decomposition, which is parsed into reconstructed components with clear periodic characteristics and enhanced stationarity. This hierarchical processing mechanism effectively addresses the error accumulation effect caused by high-frequency direct prediction and the challenges posed by non-stationary signals to the adaptability of the prediction model, thereby systematically improving the accuracy of time series prediction.

[0031] like Figure 2As shown, step S3 includes: S31 In order to improve the long-term dependence of the recurrent neural network, a gate mechanism is introduced to control the speed of information storage, including selectively adding new information and selectively forgetting previously stored information. LSTM mainly controls the transmission of data through three gates, namely, forget ; , in, It is the network structure parameter that needs to be updated in the LSTM network. Represents the long-term memory information of the time series, activation function Map the output to Interval and On the interval.

[0032] The S32 uses a BiLSTM architecture consisting of two LSTM networks, one forward and one reverse, to perform bidirectional predictions based on the data's changing patterns. Compared to LSTM, BiLSTM is more advantageous for extracting information features from complex power data used in photovoltaic power forecasting without increasing the data volume required. Therefore, using BiLSTM for power forecasting can improve model prediction accuracy. The S33DBO algorithm constructs a multimodal interaction model by analyzing five typical behaviors of dung beetle populations, including ball-rolling navigation, courtship dance, food optimization, competitive stealing, and reproductive migration. The specific steps are as follows: (1) Ball-rolling behavior. It enables dung beetles to move in the search space along the specified direction, which can be expressed as: ; Where: Indicates the current iteration number; Indicates the location information of the dung beetle; is the deflection coefficient, and its value range is ; The value range is is the natural coefficient, Indicates no deviation. Departure from the original direction; represents the global worst position; Indicates changes in light intensity.

[0033] (2) Dancing behavior. This process simulates the strategy adopted by dung beetles when facing obstacles to find new feasible routes, which can be expressed as: ; Where, .

[0034] (3) Reproduction behavior. This is a boundary selection scheme and can be expressed as: ; Where, Indicates the current local optimal position; 、 Represent the lower and upper bounds of the optimization problem; , Represents the upper bound of the spawning area. The position of the egg is also continuously updated during the iteration process, and its update formula is as follows: ; Where, .

[0035] (4) Foraging behavior.

[0036] ; Where, Indicates the upper bound of the optimal foraging area. The position of the dung beetle is updated: ; Where, is a random number that follows a normal distribution, A random vector of .

[0037] (5) Stealing behavior. The algorithm can avoid falling into the local optimal solution and can be expressed as: ; Where, is a constant; is a random vector that follows a normal distribution.

[0038] like Figure 3 As shown in Figure 3, S34: DBO algorithm for BiLSTM hyperparameter optimization process: (1) Initialize various parameters in the DBO optimization algorithm, including the number of iterations, dung beetle scale, upper and lower limits of the optimization variable size, maximum number of iterations, etc. (2) Input the data set, divide it into training set and test set, calculate the individual fitness of the DBO algorithm, and save the optimal individual and the optimal position; (3) If the optimal solution can be determined when the termination conditions are met, the program exits; otherwise, it returns to step (2); (4) The optimized parameters are reassigned to the BiLSTM network, and the network is trained to obtain the optimized photovoltaic point prediction results; the parameters of the LSTM network are optimized based on the DBO algorithm, and finally the distributed photovoltaic point prediction values ​​of the two components are obtained.

[0039] The kernel density estimation (KDE) method includes step S41, and the mathematical expression of the kernel density value is as follows: , Kernel function; ; S42. The formula of the kernel function is as follows: ; x is the sample data, and h is the bandwidth parameter of the kernel function.

[0040] S43, superposition prediction interval: The error probability density function of the oscillation component is calculated by the kernel density estimation method, and its integral is obtained Probability distribution function. At the significance level of If the error sample falls into the error interval The probability is not less than , then the interval It is called confidence Confidence interval of .

[0041] The error confidence interval of the oscillation component is superimposed with the point prediction values ​​of the trend component and the oscillation component to obtain the final prediction interval. The calculation formula is as follows: ; ; in, is the upper limit of the prediction interval, is the lower limit of the prediction interval. is the point forecast value of the superposition of trend component and oscillation component, For distributed photovoltaic rated power, the application of KDE method in prediction interval calculation can provide more accurate and reliable prediction results.

[0042] The working principle of the present invention is as follows: first, the photovoltaic power sequence is decomposed into multiple modal components by the complete ensemble empirical mode decomposition with adaptive noise (CEEMDAN) method; secondly, the high-frequency non-stationary components obtained by the primary decomposition are subjected to secondary decomposition to obtain secondary components, and the sample entropy method is used to reconstruct all components into trend components and oscillation components to quantify the complexity of each component. This indicator is positively correlated with the randomness of the sequence, and the larger the entropy value, the more irregular the signal; then, the disturbance component with the largest entropy value is selected for secondary modal decomposition, and it is parsed into reconstructed components with clear periodic characteristics and enhanced stationarity. This can eliminate the problem of inaccurate components caused by significant high-frequency non-stationary noise caused by the generation of high-order intrinsic mode functions by DEEMDAN, and the secondary modal decomposition is performed only on the disturbance component with the largest entropy value. The calculation method is simple and does not require a large number of similar algorithms for calculation and determination. Then, the bidirectional long short-term memory network (BiLSTM) is optimized by the Dung Beetle Optimization (DBO) algorithm. The parameters are optimized, and finally the distributed photovoltaic point prediction values ​​of the two components are obtained; finally, the probability density estimation of the point prediction error of the oscillation component is performed through the kernel density estimation method (KDE), and the point prediction values ​​are superimposed to obtain the overall prediction interval result. The bidirectional long short-term memory network can associate the output value with the positive and reverse conditions of the input value, making the optimal parameters more accurate. At the same time, after obtaining the distributed photovoltaic point prediction value, it is necessary to determine the probability density estimation of the point prediction error through the kernel density estimation method to obtain the estimated value. The final estimated value is limited by the area with the largest probability, and the prediction interval is narrowed. At the same time, the accuracy of the entire interval can also be made higher, which can provide new ideas for the prediction of distributed photovoltaic output interval. The obtained prediction results have important engineering practical value for ensuring the safe and stable operation of the power grid.

Claims

1. The photovoltaic output interval prediction method based on CEEMDAN-DBO-BiLSTM is characterized by: The steps include: S1 obtains photovoltaic data; and based on the nonlinear and non-stationary characteristics of photovoltaic historical data, decomposes the photovoltaic power series into two or more modal components using the adaptive noise complete set empirical mode decomposition method, the modal components including high-frequency non-stationary components; S2 uses the sample entropy method to perform secondary modal decomposition on the high-frequency non-stationary components obtained by the primary decomposition, and the secondary components are reconstructed into trend components and oscillation components; The sample entropy method quantifies the complexity of each high-frequency non-stationary component and selects the disturbance component with the largest entropy value for secondary mode decomposition; S3 optimizes the parameters of the bidirectional long short-term memory network based on the dung beetle optimization algorithm, and finally obtains the distributed photovoltaic point prediction values ​​of the two components; S4 performs probability density estimation on the point prediction error of the oscillation component based on the kernel density estimation method to obtain an estimated value, and superimposes the distributed photovoltaic point prediction value of the two components with the estimated value to obtain the overall prediction interval result.

2. The photovoltaic output interval prediction method based on CEEMDAN-DBO-BiLSTM according to claim 1 is characterized by: The steps of decomposing a photovoltaic power sequence into two or more modal components using an adaptive noise complete set empirical mode decomposition method include: S11: Add to The Gaussian white noise with sub-mean of 0 is constructed The experimental ,in is the preset Gaussian white noise weight coefficient; Gaussian white noise generated during processing; S12 pair sequence Perform EMD decomposition to obtain the first modal component IMF1 and take its mean as the first IMF obtained by the adaptive noise complete set empirical mode decomposition. ,in, represents the first modal component obtained by the adaptive noise complete set empirical mode decomposition; Represents the residual signal after the first decomposition.

3. The photovoltaic output interval prediction method based on CEEMDAN-DBO-BiLSTM according to claim 2 is characterized by: S12 and later also include: The first After adding specific noise to the residual signal, continue to perform EMD decomposition. , ;in Indicates the first IMF components; Denotes the adaptive noise complete set empirical mode decomposition method for the The weight coefficient of the noise added to the residual signal of the stage; S14 If the EMD stop condition is met, If it is a monotonic signal, the iteration stops and the decomposition of the adaptive noise complete set empirical mode decomposition method algorithm ends.

4. The photovoltaic output interval prediction method based on CEEMDAN-DBO-BiLSTM according to claim 1 is characterized by: The steps of reconstructing all components into trend components and oscillation components using the sample entropy method include: S21 optimizes the bandwidth constraint through variational mode decomposition (VMD), which can decompose high-frequency non-stationary components into more stable modal components. The VMD constrained variational model constructed in step S21 is as follows: ； are the submodal component set and its corresponding center frequency set respectively; is the pre-set or estimated number of modes; is the partial derivative operator; * is the convolution operator; is the unit pulse function; represents an imaginary number; is the signal to be decomposed; S22 adds penalty items , Lagrange multiplication operator , transformed into an unconstrained variational problem: ； The alternating direction multiplication method is used to Optimize and output the decomposed components when the required accuracy is achieved; S23 performs a sample entropy operation on the decomposed components that have reached the required accuracy output.

5. The photovoltaic output interval prediction method based on CEEMDAN-DBO-BiLSTM according to claim 1 is characterized by: Step S3 includes: S31LSTM mainly controls the transmission of data through three gates, namely the forget gate, input gate, and output gate; S32 uses a bidirectional long short-term memory network composed of two LSTM networks, which are forward and reverse. It can make bidirectional predictions based on the changing patterns of data and use the bidirectional long short-term memory network to predict photovoltaic power. The S33 DBO algorithm constructs a multimodal interaction model by analyzing the typical behaviors of dung beetle populations, such as ball navigation, courtship dance, food optimization, competitive stealing, and reproductive migration. S34 DBO algorithm for bidirectional long short-term memory network hyperparameter optimization.

6. The photovoltaic output interval prediction method based on CEEMDAN-DBO-BiLSTM according to claim 1 is characterized by: Step S34 includes: (1) Initialize various parameters in the DBO optimization algorithm, including the number of iterations, dung beetle size, upper and lower limits of the optimization variable size, and the maximum number of iterations; (2) Input the data set, divide it into training set and test set, calculate the individual fitness of the DBO algorithm, and save the optimal individual and the optimal position; (3) If the termination condition is met, the optimal solution can be determined and the program exits; otherwise, return to step (2); (4) The optimized parameters are reassigned to the bidirectional long short-term memory network, and the network is trained to obtain the optimized photovoltaic point prediction results.

7. The photovoltaic output interval prediction method based on CEEMDAN-DBO-BiLSTM according to claim 1 is characterized by: The kernel density estimation method in step S4 includes: S41 determining a kernel density formula; ,in, is bandwidth; is the total number of quantiles; is a dataset consisting of conditional quantiles; is the kernel function; The bandwidth is The kernel density expression when ; S42. The formula of the kernel function is as follows: ; Where x is the sample data, h is the bandwidth parameter of the kernel function; S43, superposition prediction interval: The error probability density function of the oscillation component is calculated by the kernel density estimation method, and its integral is obtained Probability distribution function; at the significance level of If the error sample falls into the error interval The probability is not less than , then the interval It is called confidence The confidence interval of the oscillation component is superimposed with the point prediction value of the trend component and the oscillation component to obtain the final The prediction interval is calculated as follows: ； ； in, is the upper limit of the prediction interval, is the lower limit of the prediction interval. is the point forecast value of the superposition of trend component and oscillation component, is the photovoltaic rated power.

Images