Short-term prediction of photovoltaic power generation based on ICEEMDAN-SE-GAPSO-LSTM

This study proposes an improved adaptive noise-complete ensemble empirical mode decomposition (ICEEMDAN), sample entropy (SE), and genetic particle swarm optimization (GAPSO) algorithm-based short-term solar power forecast model. The photovoltaic power data is first decomposed using the ICEEMDAN algorithm to produce a number of eigenmode components with various properties. The sample entropy of each component is then calculated, and the components with similar sample entropy values are combined to create new modal components. These new modal components are then input into the GAPSO-optimized LSTM model for prediction, and the prediction results of each component are summed and reconstructed to produce the final prediction. The results completely validate the efficacy and dependability of the suggested method, using the measured data from a photovoltaic power station in Xinjiang, China as an example.


Introduction
Under the accelerating wave of renewable energy revolution around the world, photovoltaic (PV) power generation, as one of the mature and widely used renewable energy technologies, is gradually replacing traditional fossil energy sources as the main source of electricity [1][2].Short-term PV power forecasting can help power system operators carry out rational scheduling and optimize the balance between power supply and demand.By accurately forecasting PV power generation, grid scheduling can be rationally arranged to ensure stable power supply from the PV power system while optimizing the operational efficiency of the grid.Accurate power forecasting has a significant impact on market transactions and power pricing and can help power market participants to rationally formulate strategies for purchasing and selling power, as well as determining power pricing, thereby maximizing market benefits [3].Accurate forecasting results can assist governments and energy administrators in evaluating the potential and viability of photovoltaic (PV) power generation and formulating practical energy policies and development plans.Short-term PV power prediction also has guiding significance for the maintenance and operation management of PV equipment.
Artificial intelligence techniques have steadily developed to further increase the precision and dependability of PV prediction.ICEEMDAN [4] can adaptively adjust the noise level and improve the quality of the decomposition results as well as its accuracy and stability by adding a set of kth-order IMFs in a special white noise and thus reducing the residual noise compared to the CEEMDAN decomposition.The LSTM model introduces a gating mechanism that can selectively remember or forget the information in the input sequence.When designing the LSTM model for PV power prediction, intelligent algorithms are used to optimize its parameters.Optimization mechanisms such as Particle Swarm Algorithm (PSO) [5], Simulated Annealing Algorithm (SA) [6], and Bayesian Optimization (BO) [7] are usually used to optimize the model parameters of LSTM.
In summary, this paper proposes a short-term PV power prediction model based on ICCEEMDAN-SE-GAPSO-LSTM. Taking the measured data of a PV power plant in the Xinjiang region as an example, and temperature, humidity, barometric pressure, total radiation, and direct radiation as the influencing factors, the effectiveness of the model proposed in this paper is verified through simulation.

ICEEMDAN decomposition
ICEEMDAN decomposition [4] is a further pair of improvements based on the CEEMDAN algorithm.By adding a set of kth-order IMFs in a special white noise, the adaptive nature of the IMFs can be utilized to adaptively match the original signals to reduce the noise component [8], which can better deal with modal aliasing compared to the CEEMDAN decomposition, and ICEEMDAN decomposition can adaptively adjust the noise level to inhibit the noise influence on decomposition results.The steps of the ICEEMDA algorithm are as follows: We define ( ) where  is the signal-to-noise ratio; N is the amount of white noise added.
(2) The modal components of the first stage (k=1) can be derived.
( ) where k r can be denoted as

Sample entropy theory
Since sample entropy [9] has better adaptability to nonlinear time series, captures complex nonlinear relationships, and is robust to noise and outliers, it performs better in predicting complex nonlinear time series.Therefore, to reduce the computational complexity of the prediction process caused by the more components generated by the decomposition of PV data by the ICEEMDAN algorithm, this paper introduces the sample entropy algorithm to reconstruct the components generated by the decomposition.

Long-and short-term memory networks
LSTM [10] is a recurrent neural network structure commonly used in time series prediction, and the LSTM model is more sensitive to the parameters.If the hyperparameters cannot be effectively determined, the prediction accuracy will be greatly reduced.Due to the intermittent and fluctuating nature of PV power generation and its susceptibility to external factors, additional preprocessing and feature engineering steps may be required to effectively incorporate them into the model, and these components will interfere with the prediction performance of PV power.Therefore, the parameters of the LSTM model need to be optimized so that the data for PV prediction will be more accurate.

GA-PSO algorithm
The Genetic Particle Swarm algorithm [11] takes the best parts of both the Genetic Algorithm [12] and the Particle Swarm algorithm and puts them together.The genetic algorithm may converge more slowly at the beginning because its selection operations are not always correct.The genetic particle swarm algorithm, on the other hand, uses the local search features of the particle swarm algorithm to do a better job of finding the best solution in the area.It can converge to a solution faster at the beginning, which makes the initial convergence speed better.It is easy for the particle swarm algorithm to find local optimal solutions because particles share information [13].

Model construction
Figure 1 shows the short-term PV power prediction model based on ICEEMDAN-SE-GAPSO-LSTM established in this paper, and its specific steps are as follows: (1) We decompose the raw PV power data into corresponding IMF components and residuals by ICEEMDAN decomposition.
(2) We determine the sample entropy for every component and then rearrange the components into new components based on sample entropy values that are similar.
(3) We optimize the relevant parameters of the LSTM model by the GAPSO algorithm and establish the prediction model of GAPSO-LSTM for each restructured component.
(4) We use the GAPSO-LSTM model to forecast each restructured component.The final prediction results are achieved by superimposing the prediction results.

Example analysis
The actual PV data from a place in Xinjiang, China, is used for analysis, and the PV power generation data from July 19, 2019, to September 27, 2019, is selected.The PV power generation period from 7 a.m. to 9 p.m. is selected, and the data interval is 15 minutes, so the final period is 71 days, with a total of 4, 047 data points.The data before September 26 is selected as the test set, and September 26 is the validation set.ICEEMDAN is used to decompose the PV data.In the paper, the noise standard deviation and integration number of ICEEMDAN decomposition are set to 0.2 and 50.The decomposition results are shown in Figure 2. A total of 11 components are obtained, including 10 IMFs and one residual component.The sample entropy of each component is computed and the outcomes are presented in Figure 3.It is observed that IMF3 exhibits the highest sample entropy value, suggesting a greater level of complexity.Conversely, IMF11(res) displays the lowest sample entropy value, indicating a lower level of complexity.This paper categorizes the components based on their sample entropy values.Components with a sample entropy value greater than 0.35 are grouped, components with a sample entropy value ranging from 0.1 to 0.35 are grouped, and components with a sample entropy value lower than 0.1 are grouped.The reconstruction results for each component are presented in Table 1, while the reorganized components are depicted in Figure 4.The GAPSO-LSTM model is utilized for the prediction of individual recombination components.The ultimate prediction results are obtained by combining and synthesizing the expected outcomes of each recombination component.Models for photovoltaic (PV) data from ICEEMDAN-SE-PSO-LSTM, GAPSO-LSTM, PSO-LSTM, and LSTM are utilized for assessing the proposed model's predictive ability.The comparison results are presented in Figure 5.These evaluation metrics are summarized in Table 2.In contrast to alternative models, the model presented in this research has the greatest magnitude and superior congruence with the source data.Compared with ICEEMDAN-SE-PSO-LSTM, GAPSO-LSTM, PSO-LSTM, and LSTM models, the MAE of the proposed model is reduced by 0.91, 1.067, 1.551, and 1.936, respectively; and the RMSE is reduced by 1.26, 1.333, 2.112, and 2.47, respectively.In conclusion, the visualization graphs and index comparisons can be viewed.By using charts, graphs, and comparisons of different indices, it is readily apparent that the suggested model outperforms the standard model for predicting PV over the near term.


 as the operator that creates the local mean of the applied signal and define the operator ( ) k E  as the kth modal component found by using the EMD method of decomposition.We set i as white Gaussian noise with a mean of 0 in a normal distribution.ICEEMDAN's kth modal component, which is k IMF , is based on the EMD computational process.decomposition implemented by EMD computation to obtain the first residue.We add a special set of white noise ( ) i E  of the sequence ( ) i L , and calculate the local mean value of the sequence signal obtain the first residue.

Figure 3 .
Figure 3. Plot of sample entropy results for each component.

Figure 4 .
Figure 4. Graph of the results of each recombination component.

Figure 5 .
Figure 5.Comparison of PV power prediction results.

Table 1 .
Table of reorganization details for each component.

Table 2 .
Model evaluation metrics.In this paper, a new short-term PV power prediction model called ICEEMDAN-SE-GAPSO-LSTM is suggested.First, ICEEMDAN is utilized to identify the eigenmode functions at different scales, and then sequences of samples with similar entropy values are put back together.Then, GAPSO is used to solve the problem of how hard it is to figure out LSTM's structural factors.The GAPSO-LSTM model is used to predict each reconstructed component, and the prediction results are added together to get the final PV data prediction results.These results are simulated and checked with the actual PV power generation data of a certain place in Xinjiang, China, as an example.The results show that the model proposed in this paper has a better prediction effect than the traditional prediction model, and it performs well for short-term PV power generation.