The power output state of photovoltaic power generation is affected by the earth's rotation and solar radiation intensity. On the one hand, its output sequence has daily periodicity; on the other hand, it has discrete randomness. With the development of new energy economy, the proportion of photovoltaic energy increased accordingly. In order to solve the problem of improving the energy conversion efficiency in the grid-connected optical network and ensure the stability of photovoltaic power generation, this paper proposes the short-term prediction of photovoltaic power generation based on the improved multi-scale permutation entropy, local mean decomposition and singular spectrum analysis algorithm. Firstly, taking the power output per unit day as the research object, the multi-scale permutation entropy is used to calculate the eigenvectors under different weather conditions, and the cluster analysis is used to reconstruct the historical power generation under typical weather rainy and snowy, sunny, abrupt, cloudy. Then, local mean decomposition (LMD) is used to decompose the output sequence, so as to extract more detail components of the reconstructed output sequence. Finally, combined with the weather forecast of the Meteorological Bureau for the next day, the singular spectrum analysis algorithm is used to predict the photovoltaic classification of the recombination decomposition sequence under typical weather. Through the verification and analysis of examples, the hierarchical prediction experiments of reconstructed and non-reconstructed output sequences are compared. The results show that the algorithm proposed in this paper is effective in realizing the short-term prediction of photovoltaic generator, and has the advantages of simple structure and high prediction accuracy.

With the development of photovoltaic power generation, the proportion of solar energy in renewable energy is increasing, and the installed capacity of photovoltaic power stations is also increasing [

At this stage, the power prediction of photovoltaic power generation mainly includes direct prediction and indirect prediction [

In direct prediction, literature [

In indirect prediction, literature [

Based on the above analysis, to eliminate the influence of weather conditions, in addition to the indirect prediction method, this paper also improves the direct prediction method and proposes the short-term prediction of photovoltaic power generation based on LMD permutation entropy and singular spectrum analysis. Firstly, multi-scale permutation entropy was used to measure the characteristics of PV power series under different weather types, and the historical power series were reconstructed in daily units. Secondly, the hierarchical prediction method is used for the recombination sequence, and the LMD algorithm is used to decompose it to extract more frequency components. Finally, the singular spectrum analysis algorithm is used to predict the decomposed components, and the prediction results of each component are reconstructed. The validity of the proposed algorithm is verified by an example under different prediction algorithms, and the prediction accuracy is improved.

The main innovation points of this paper are as follows: 1) To eliminate the influence of weather factors, the multi-scale permutation entropy is used to classify and process the photovoltaic power generation sequence under different meteorological conditions. 2) After reassembling the reconstructed sequence, the LMD algorithm is used to extract the detailed components. 3) Use the singular spectrum analysis algorithm to achieve the hierarchical prediction of photovoltaic power generation.

The calculation of output current and output power of photovoltaic power generation on the unit array is shown in

In _{L} is the photocurrent determined by the solar radiation intensity; _{D} is the diode current determined by the number of photovoltaic cells and _{S} is determined by the resistance on the PV panel, _{S} and _{SH} are the PV voltage, series, and parallel resistance, respectively;

According to the above analysis, photovoltaic power generation status is mainly determined by meteorological conditions: temperature, sunshine amount, solar radiation angle, and weather conditions. In this paper, the PV output power between sunrise 6:00 and sunset 21:00 is firstly counted. According to the historical data of a power plant, taking the meteorological data of the China Meteorological Bureau on rainy and snowy days, sunny days, abrupt days, and cloudy days as examples, its output power curve is shown in

According to the output power curves under different weather conditions, the output power is affected by the coverage of rain and snow on snowy days, and the PV output power is the lowest. In lightning weather, influenced by meteorological mutation, photovoltaic power output randomness is strong; under the influence of solar radiation intensity, the PV output power is lower on cloudy days and sunny days.

Thus, on the other hand, the state of photovoltaic output power is determined by meteorological conditions, and on the other hand, its mechanical structure. To realize the short-term prediction of its output power, the first thing to be solved is to quantitatively analyze the output power under different meteorological conditions, to make a more accurate short-term prediction of photovoltaic power generation.

As a signal processing method, local mean decomposition has the advantage of processing nonlinear and non-stationary signal sequences [

The local mean decomposition algorithm uses the extreme points of time series to process the moving average value and decomposes the original signal sequence

_{i}(

Assume that the local extreme point of the original time series _{j}, where _{i} of two adjacent extreme points is taken as the window, and the local mean component _{ik} is formed after sliding and smoothing processing. The corresponding _{i} is the estimated value of the envelope, and the envelope decomposition component _{ik} is formed after sliding smoothing.

In the decomposition process, _{ik}(_{ik}(

Multi-scale permutation entropy is a measure of time sequence complexity index [

Step1: Coarse granulation treatment

Taking the scale factor

When the number of sampling points is insufficient, such as

1) In the coarse granulation process of the one-dimensional signal sequence, the number of samples decreases, and the feature representation is not sufficient;

2) Calculating the mean value under the scale factor may lead to decreased sensitivity to signal sequence.

Therefore, in this paper, the coarse granulation process is improved progressively. Assuming

Firstly, the mean values of

Then, the above four coarse-graining values are taken as the first value of the improved coarse-graining sequence, then the sequence of the second value of the first group of sequences after corresponding improvement is

Finally, the operations of Step2, Step3, and Step4 are carried out on the four groups of coarse granulation respectively, and then the mean values of the four groups are taken to complete the improvement of the multi-scale permutation entropy.

Step2: Phase space reconstruction

Assuming that the time delay parameter is

Step3: Calculation of permutation entropy

Each row of the reconstructed matrix is arranged in descending order, and the index column number of the column where the vector element value resides can be expressed as:

The probability value of _{l}. After statistics, the calculation result of multi-scale permutation entropy is shown in

Step4: Standardized treatment

To realize the normalization and standardization of the entropy results, the entropy results of Step3 are calculated as shown in

The above entropy calculation process is mainly used to calculate the characteristic parameters of time series. As a kind of characteristic quantity, multi-scale permutation entropy is conducive to the establishment of signal sequence measurement standards on the one hand and improves the accuracy of subsequent signal sequence processing when used as a correction parameter on the other hand.

SSA is a regression prediction algorithm for discrete signal sequences, which is suitable for processing time series of nonlinear and potential structures [

Assumption: non-zero original signal sequence ^{*} is shown in

Step1: Establish the trajectory matrix

Take the integer

Step2: Singular value decomposition

Firstly,

Then,

Finally,

Step3: Grouping of singular decomposition matrices

The set {1,2,…,_{I} are shown in

Step4: Recombination of the one-dimensional signal sequence

Signal sequence recombination mainly adopts the diagonal averaging method to recombine each grouping matrix _{1},…,_{N}, the specific calculation steps are shown in

^{*} = min(^{*} = max(

Step5: Prediction of one-dimensional signal

Two algorithms, recursion, and matrix are mainly used for prediction using singular value decomposition results. The recursive algorithm with a simpler algorithm is adopted in this paper, and its main ways are as follows:

_{i} is an orthonormal vector in singular value decomposition, and _{i}.

The singular spectrum decomposition algorithm uses a singular value matrix to decompose the original signal sequence into different components from three directions of trend, vibration component, and an unstable factor. SSA algorithm is not only suitable for nonlinear and discrete time series but also has the advantage of not requiring specific parameter models and stability conditions.

In the process of photovoltaic power generation, its output power is unstable due to the influence of uncertain factors. To better realize photovoltaic grid connection and improve power supply quality, this paper proposes a short-term prediction algorithm for photovoltaic power generation based on LMD permutation entropy and singular spectrum analysis. The specific process is shown in

1) Photovoltaic power generation is affected by the rotation of the earth, and its output power has a 24-h periodicity. To eliminate the influence of weather factors on the output power of photovoltaic power generation, the meteorological conditions can be roughly divided into sunny weather (sufficient sunshine), rainy and snowy weather (less sunshine), cloudy day (insufficient sunshine), and abrupt weather (strong randomness of sunshine) according to the monitoring data of the Meteorological Bureau and the permutation entropy characteristic analysis.

According to the above analysis, to better realize the short-term prediction of photovoltaic power generation, firstly, the historical power curve of photovoltaic power generation is classified according to four meteorological types. Then the output power curves of _{1}, _{2}, _{3,} and _{4} under four meteorological conditions are formed. Finally, the accuracy of the short-term forecast can be improved according to different meteorological conditions.

2) _{1}, _{2}, _{3}, and _{4} are selected as historical output power curves under different meteorological types according to meteorological prediction and clustering results under the next one-day cycle. To realize the classification prediction of photovoltaic power generation, the LMD algorithm is used to process it, and

3) Each decomposed component is analyzed and predicted by a singular spectrum, and the hierarchical prediction results of IFM(_{1}), IFM(_{2}), …, IFM(_{n}) are obtained. Then, The multi-scale permutation entropy is used to construct photovoltaic series under different weather types, and SSA prediction is carried out for them, and the error sequence _{1} is obtained by comparing with the initial sequence; The prediction sequence is predicted again, and the error sequence _{2} is obtained by comparing with the prediction sequence; By analogy, a complete error sequence is obtained to correct the prediction results of different components, thus reducing the error. Finally, through the reconstruction under the LMD algorithm, the photovoltaic power generation in the corresponding next-day cycle is obtained, to realize the short-term prediction of power generation.

According to the analysis in

Photovoltaic power generation cycle within 24 h, taking 15 min as the sampling point, the number of sampling points from 6:00 to 21:00 is

As shown in

Firstly, the action power curve is divided into a sample set and a test set. The sample set includes Group 1, Group 2, …, Group

Taking a sunny day as an example, the reconstituted output sequence is shown in

It can be seen from the generation power diagram in

Given the power sequence reconstruction in the above section, the LMD algorithm has the advantage of self-adaptation and decomposition, which is conducive to the realization of hierarchical prediction. The decomposition process of LMD is shown in Algorithm 1. Taking the reconstructed power series under sunny days as an example, the results after LMD processing are shown in

As shown in _{4}), the distribution of extreme points is insufficient and meets the conditions for termination of decomposition. Under the same weather type, the center frequencies of different decomposition components are different. In different weather types, the center frequency of the same decomposition component is also different. After the original signal is processed by LMD, it is beneficial to achieve hierarchical prediction and improve the prediction accuracy under the condition that the properties of the original signal sequence remain unchanged.

If the original output power is simply used as the prediction input sequence, the influence of weather factors is not considered for hierarchical prediction. Taking the historical power generation series of a power plant as an example, this paper reorganizes the photovoltaic power generation series under sunny weather in spring and uses the reconstructed 16-day historical power data as the training set. According to the weather forecast of the Meteorological Bureau, the singular spectrum analysis algorithm is used to make short-term forecasts for the next day, and the prediction result is shown in

According to the graded prediction results in

Reconstruction-Classification short-term prediction is to reconstruct the original sequence according to the weather type, then use LMD classification prediction, and finally recombine the different decomposition components to achieve short-term prediction of photovoltaic power generation. When the weather of the day is predicted to be sunny by the Meteorological Bureau, the recombination sequence under sunny weather conditions is used as the historical power. When the weather of the predicted day is rain and snow, the recombination sequence under typical rain and snow weather conditions is used as the historical power. And the same goes for four typical weather conditions.

Taking the reconstructed sequence in typical sunny weather in

Compared with the predicted results in

To further prove the advantages of singular spectrum analysis algorithm in photovoltaic power generation sequence prediction, this paper selects multivariable support vector (SSVM) [

Prediction algorithm | MRE/% | RMSE/% |
---|---|---|

SSVM | 5.12 | 6.15 |

ELM | 4.56 | 8.64 |

BP | 3.84 | 5.14 |

SSA | 1.51 | 3.17 |

Through the above analysis, the effectiveness of using multi-scale permutation entropy to reconstruct the output sequence is proved, the influence of weather and other uncertain factors is eliminated, the advantage of using the SSA algorithm to deal with the one-dimensional discrete sequence is improved, and the prediction accuracy of photovoltaic power generation is improved.

Given the unstable phenomenon of PV power output state, this paper proposes the short-term prediction of PV power based on LMD permutation entropy and singular spectrum analysis, and the following conclusions are drawn through case demonstration:

1) Using multi-scale permutation entropy as the state characteristic value of output power under different meteorological conditions, the historical output sequence is reconstructed to eliminate the influence of meteorological conditions on the forecast sequence.

2) LMD algorithm has the advantage of self-adaptation in signal processing and achieves hierarchical prediction for the reconstructed sequence.

3) Using the advantage of singular spectrum analysis in processing discrete signal sequences, the short-term power prediction is realized by the reconstructed sequence after the hierarchical prediction of decomposed components.

Finally, an example is used to verify the short-term prediction of pure signal sequences considering weather conditions. Then, the short-term prediction of photovoltaic sequences reconstructed under the multi-scale permutation entropy algorithm under different meteorological conditions is compared. The experimental results prove the improvement of RMSE and MES of the proposed algorithm in photovoltaic prediction. Compared with the experimental results under different prediction algorithms, it is proved that singular spectrum decomposition is more beneficial to realizing photovoltaic grid connection.

The author is very grateful for the experimental platform provided by Zhengzhou University of Railway Engineering.

The author received no specific funding for this study.

The author declares that they have no conflicts of interest to report regarding the present study.