Short-term PV power prediction study based on FCM-BiGRU

Addressing the issue of large prediction error in short-term load prediction of existing neural network models, this study proposes a short-term load forecasting approach that combines fuzzy C-mean clustering and a two-way gated recurrent neural network model. Fuzzy C-mean clustering is first applied to cluster the original data into three typical days, and the grouped data are trained using a bidirectional gated recurrent neural network model for load prediction. The conclusive experiment demonstrates that the proposed approach introduced in this study exhibits a high prediction accuracy in the context of short-term photovoltaic output forecasting, and there is also a substantial error reduction compared with the existing neural network methods.


Introduction
The world's energy is in urgent need of transformation, and in the context of "carbon peak, carbon neutral", the development of renewable energy represented by wind and light energy is promising.Continuous in-depth construction of energy Internet, distributed new energy, energy storage, electric vehicles, demand response load and other large resources will be connected to the grid, which will bring huge challenges to the efficient use of flexible resources and the traditional power network power supply security and stable operation mechanism [1].Currently, the literature has employed the Sparrow Algorithm (SSA)-BP neural network method to achieve short-term power forecasting of PV systems, through the Sparrow Algorithm to the BP neural network's power threshold for fast optimization, and then improve the network's prediction accuracy [2].Yang et al. [3] used the combination algorithm of wavelet decomposition (WD) and error backpropagation algorithm (Backpropagation, BP) neural network to accomplish shortterm forecasting for solar electricity generation facilities.Xu et al. [4] introduced the KPCA-K-means algorithm to reduce the dimension classification of original data, so as to train GRU model parameters of different categories to improve prediction accuracy.
Although the above studies have made improvement strategies for short-term solar power prediction models, most of them tend to fall into local optimization or use historical data insufficiently.In this paper, we propose to use a fuzzy C-mean (FCM) clustering algorithm to perform unsupervised clustering of the data, classify the PV power generation according to the typical data, and finally fit the data by using a bi-directional gated recurrent unit (BiGRU), which further improves the accuracy of the shortterm PV electricity generation forecasting.

Analysis of PV output power influencing factors
PV generation is impacted by different weather elements [5].When the positive distribution is not satisfied among the data, the Kendall rank correlation coefficient, employed in the non-parametric statistical method, is utilized to quantify the level of association among the elements.The statistical method can more intuitively demonstrate the impact magnitude of different weather factors on the magnitude of photovoltaic electricity generation through data [6].The expression of Kendall's rank correlation coefficient is: ( ) where X and Y represent the count of matching pairs and divergent pairs, in that order; and N represents the quantity of samples; and ( ) NN − represents the quantity of two-by-two combinations of all samples.
The heat map of correlation coefficients is shown in Figure 1.Meteorological variables with a correlation sanhedrin exceeding 0.4 are identified as the primary variables influencing photovoltaic output in this study.

FCM algorithm
When making PV output forecasts, scholars usually aggregate data with similar characteristics into one category [7].FCM clustering algorithm introduces fuzzy degrees on the basis of a traditional clustering algorithm.Compared with the traditional K-means hard clustering method, FCM has a more flexible clustering effect.The update of the affiliation matrix is shown in Equation ( 2

Clustering feature selection
The mean, standard deviation, maximum, number of peaks and valleys, variability ratio, peakness, and asymmetry of the total solar radiation variables with the highest power correlation coefficients were selected as the clustering features based on the heat map [8].The number of cluster centers K not only needs to be determined manually, but also needs to determine whether the K value is selected properly [9], in this paper, the contour coefficient method is used for clustering evaluation.Table 1 presents the contour coefficients related to various K values.The silhouette coefficient takes on a value between -1 and 1, with a score of 1 denoting accurate classification to the correct swarm and a score of 0 denoting erroneous classification to an inappropriate swarm.When K=3, the corresponding profile coefficient is 0.5968.determine the value of K for the number of cluster centers selected in this paper is 3.  [10].LSTM controls cell state through three different types of gated units to control, update and transfer state information.The gating unit regulates the data flow, and the memory state facilitates retaining information in the temporal dimension for an extended period, thus solving the problem of gradient parameter information superposition in recurrent neural networks.At the same time, increasing the number of parameters requires more training time and computational resources [11].The GRU comprises constructed solely with reset and update gates.which addresses the issue of gradient blowing up or gradient vanishing during the processing of extended time series information by RNN, and retains the ability to process time series [12].BiGRU neural network discovers the intrinsic relationship and regularity of historical, current and future time point data by bidirectional analysis of time series data, The computational equation for the GRU neural network is: ( ) ( ) ( ) ( ) ( )  E coefficient of determination is used as the evaluation index of the model fitting effect.The computation of the aforementioned evaluation metrics is as outlined: N is the total number of samples predicted.

Example analysis
In this research, data from a PV array at the Alice Spring site in 2017-2018 is used as an example of arithmetic to analyze the data.Typical daily clustering results are shown in Figure 3.The historical dataset was finally clustered into three typical day datasets, namely sunny, sunny to cloudy, and cloudy, based on the seven data characteristics of the total horizontal radiation variable with the highest power correlation.To validate the effectiveness of the proposed methodology in short-term PV power prediction, the forecasting outcomes are contrasted and visualized using RNN and GRU neural network models across three typical days, respectively.The error values of different models for each typical day are presented in Table 2.In Figure 4, a comparison is depicted illustrating the variations in power prediction outcomes for each model under diverse weather conditions.

Conclusion
• There is a large variability in PV power generation for different weather types, so weather types need to be categorized.• FCM algorithm can effectively cluster past photovoltaic generation records into typical daily datasets, and clustering data with similar characteristics can more fully mine the characteristics of historical data, and obtain the best network parameters according to various typical daily models.• Performance index errors were compared with other models under three typical days, and according to the test data and concluded that the proposed model prediction curve is correct and with the actual data fitting degree is higher, higher prediction precision.

Figure 1 .
Figure 1.Heat map of correlation coefficients.
), and the equation of the clustering center is shown in Equation (3).
affiliation of the p sample to the q cluster; pq d denotes the distance from the p sample to the centroid of the q class; pc d denotes the distance from the p sample to the centroid of the c class; m denotes the customized fuzzyness parameter; e represents the quantity of iterations; k represents the quantity of clusters; s represents the quantity of samples; p x denotes the sample points.

Figure 2 .
Figure 2. Schematic representation of the neural network structure.

w
parameter weights between the input layer and the concealed layer; are the parameter weights from one concealed layer to another concealed layer; the probability evaluation indexes of the model[14].The 2 R

Figure 3 .
Figure 3. Plot of clustering outcome across diverse typical days.

Figure 4 .
Figure 4. Contrast of PV prediction for each model under typical day.

Table 1 .
Contour coefficients corresponding to different values of K.

. Bidirectional gated recurrent unit (BiGRU) network modeling Recurrent
Neural Networks (RNNs) constantly store historical moments of information when dealing with prediction problems about time series.However, the long-term dependency problem of RNN can lead to gradient explosion or gradient vanishing problem, which makes the model training not work properly

Table 2 .
Error values of different models for each typical day.

Table 2
demonstrates that the R2 determination coefficients for all power predictions exceed 95%.When the BiGRU model is used for power prediction during sunny days, the indicator MAE undergoes a decrease of 0.042, and RMSE experiences a decrement of 0.029 relative to RNN.Under sunny conditions, the solar panels are able to capture a large amount of solar radiation and convert it into electrical energy, realizing high energy production.When it is clear to cloudy, The indicator MAE of the BiGRU model undergoes a decrease of 0.021, and RMSE witnesses a diminish of 0.017 in comparison to RNN.Compared to the GRU model, MAE is reduced by 0.021 and RMSE is reduced by 0.019.On cloudy days, the indicator MAE of the BiGRU model was reduced by 0.01 and 0.012 and the RMSE decreased by 0.012 and 0.015.