CN115115125B - Photovoltaic power interval probability prediction method based on deep learning fusion model - Google Patents

Abstract

The invention discloses a photovoltaic power interval probability prediction method based on a deep learning fusion model, which specifically comprises the following steps: firstly, obtaining a photovoltaic power and meteorological factors, performing variable correlation analysis, and determining an input variable of a prediction model; selecting a clustering variable, constructing the statistical characteristics of the clustering variable, adopting a fuzzy C-means clustering algorithm to perform similar daily clustering of the photovoltaic historical data, and performing normalization processing on the similar daily clustering; then dividing the similar daily data set into a training set and a testing set; building a QR-CNN-BiLSTM interval prediction model, training, and predicting a photovoltaic power interval; and finally, generating a photovoltaic probability prediction result on the test set. The method can well track future photovoltaic power trend, realize high accuracy measurement of photovoltaic power prediction uncertainty on the basis of meeting reliability requirements, generate a photovoltaic power prediction interval under a corresponding confidence level, and have practical application value.

Description

Photovoltaic power interval probability prediction method based on deep learning fusion model

Technical Field

The present invention belongs to the technical field of photovoltaic power generation prediction, and specifically relates to a photovoltaic power interval probability prediction method based on a deep learning fusion model.

Background technique

In recent years, environmental pollution has become increasingly serious, and the problem of shortage of non-renewable energy has become increasingly prominent. Countries around the world have begun to seek new energy development paths. my country began to implement the new energy development strategy in 2014. As an important part of energy transformation, solar energy has begun to be developed and utilized on a large scale. By the end of 2021, my country's total installed capacity of photovoltaic power generation has reached 306 million kW, and the country's new photovoltaic installed capacity in 2021 was 53 million kW. Large-scale new energy grid-connected power generation represented by photovoltaic power generation is an unstoppable development trend and outstanding feature of the new generation of power systems in the future. However, photovoltaic power generation is affected by a variety of complex environmental factors, resulting in strong random volatility, intermittency and non-stationarity in photovoltaic power generation. With the access of a high proportion of photovoltaic power generation in the power system, it will seriously threaten the safe and stable operation of the power system as an uncontrollable power source. Therefore, the study of photovoltaic power prediction technology is of great significance for my country to build a new generation of power systems and adapt them to the access of a high proportion of renewable energy. At the same time, it is of great value to build a comprehensive security defense system for the power system and realize risk control.

Existing photovoltaic power prediction technologies can be classified into deterministic prediction and uncertain prediction based on the form of prediction results. The deterministic prediction result of photovoltaic power is a single-point prediction, which has the advantage of being more intuitive, but cannot characterize the uncertainty information of photovoltaic power prediction. Uncertainty prediction can give the possible range of changes and confidence of photovoltaic power at future moments. The uncertainty prediction result can provide more comprehensive data support for power system dispatching and has more important engineering significance. From the perspective of prediction models, it can be divided into physical models and data-driven models. Physical models are established based on the characteristics of photovoltaic components, installation angles, etc., while taking into account geographical conditions and meteorological factors. The construction mechanism of physical models is complex and is currently less used. Data-driven models mainly include statistical methods and artificial intelligence algorithms. Statistical methods use curve fitting and parameter estimation to establish the relationship between photovoltaic power and its influencing factors. Commonly used methods include time series method and gray model. Artificial intelligence models are represented by neural networks and deep learning models. Artificial intelligence models have strong nonlinear data processing capabilities and are widely used in recent years.

Although many scholars have done a lot of work in the field of photovoltaic power prediction, the following problems still exist at this stage: (1) Photovoltaic power prediction is mostly concentrated on point prediction, and there is little research on interval probability prediction. (2) The reliability and sensitivity of the existing interval probability prediction model are not high, and the performance of the photovoltaic power interval probability prediction model needs to be further improved. (3) The example data used in most studies are based on 1h or 15min intervals. However, when it comes to predicting future power at 5min intervals, it will face more complex and changeable photovoltaic power fluctuations. The traditional single model will not be able to cope with this problem well, and multi-model fusion will be one of the future solutions.

Summary of the invention

The purpose of the present invention is to provide a photovoltaic power interval probability prediction method based on a deep learning fusion model, which solves the problem of inaccurate photovoltaic power interval prediction and probability prediction results.

The technical solution adopted by the present invention is a photovoltaic power interval probability prediction method based on a deep learning fusion model, which is specifically implemented according to the following steps:

Step 1: Obtain pre-processed meteorological element data and historical photovoltaic power data, perform variable correlation analysis on photovoltaic power and meteorological factors, and determine the input variables of the prediction model;

Step 2: Select clustering variables and construct statistical characteristics of clustering variables;

Step 3: Based on the clustering variables and their statistical characteristics selected in step 2, the fuzzy C-means clustering algorithm is used to cluster similar days of photovoltaic historical data to obtain a photovoltaic similar day data set;

Step 4: Use the min-max normalization method to normalize the photovoltaic similar day data set;

Step 5: Divide the similar day data sets under various weather conditions into training sets and test sets;

Step 6: Build the QR-CNN-BiLSTM interval prediction model, set the parameters of the CNN and BiLSTM models, and set the relevant parameters for model training; input the training set of the similar day data set into the model for training;

Step 7: Input the test set data into the QR-CNN-BiLSTM interval prediction model trained in step 6 to perform photovoltaic power interval prediction;

Step 8: Denormalize the interval prediction results to make them physically meaningful;

Step 9: Use the kernel density estimation algorithm optimized by cross-validation and grid search methods to generate photovoltaic probability prediction results on the test set.

The present invention is also characterized in that:

In step 1, specifically:

Step 1.1, select the pre-processed meteorological element data and photovoltaic power generation data;

The time resolution of meteorological element variables and photovoltaic power variables is 5 min, and wind speed, relative humidity, ambient temperature, total horizontal radiation, diffuse horizontal radiation, rainfall, wind direction, total oblique radiation and diffuse oblique radiation are selected as the original meteorological element data variables;

Step 1.2, the Kendall rank correlation coefficient R is used to measure the degree of correlation between multiple meteorological element variables;

Step 1.3: Select meteorological factors whose absolute value of the Kendall rank correlation coefficient R of photovoltaic power is not less than 0.5 and use them as input to the prediction model.

In step 2, specifically:

Step 2.1, select the meteorological factor variable with the highest kendall rank correlation coefficient value of photovoltaic power as the clustering variable;

Step 2.2: Select the mean, standard deviation, maximum value, number of peaks and troughs, coefficient of variation, kurtosis and skewness of the clustering variable as statistical features.

In step 3, specifically:

Step 3.1, based on the statistical characteristics of the clustering variables constructed in step 2, calculate the values of the seven statistical characteristics of the clustering variables on each day;

Step 3.2, determine the number of data clustering categories c, initialize the cluster center V _i , give the fuzzification parameter m, initialize the membership matrix U ⁽⁰⁾ , and give the termination criterion ε of the algorithm;

Step 3.3: Calculate all cluster centers of the tth iteration according to formula (5) to obtain the cluster center matrix:

Where: _uij is the membership of the i-th sample to the j-th class; _xi is the sample point; m is the membership factor; n is the number of samples; t is the number of iterations; c represents the number of clusters;

Step 3.4: Update the membership matrix U ^(t) . The calculation method is shown in formula (6):

Where: u _ij is the membership of the ith sample to the jth class; _xi is the sample point; m is the membership factor; n is the number of samples; t is the number of iterations; c is the number of clusters; d _ij is the distance from the ith sample to the center of the jth cluster;

Step 3.5, calculate ||U ^(t) -U ^(t-1) ||, and verify whether the iteration stop condition ||U ^(t) -U ^(t-1) || < ε is met. If the condition is met, stop the iteration; otherwise, continue to repeat steps 3.3 and 3.4 until the condition is met, and finally obtain a similar daily data set under various weather conditions.

In step 6, specifically:

Step 6.1: construct a feature map of the normalized similar day data set in the form of a sliding window and input it into the CNN network, and use its convolutional layer and pooling layer to extract the feature vector that characterizes the dynamic change of photovoltaic power;

Step 6.2: Convert the output feature vector into a time series and input it into the BiLSTM network;

Step 6.3: Introduce the QR model and combine it with CNN-BiLSTM. The QR model and CNN-BiLSTM model are fused in the form of loss function.

set up represents the CNN-BiLSTM point prediction model, where _Xt is the model input, i.e., the independent variable, Ω is the model parameter of CNN-BiLSTM, and _Yt is the dependent variable. is the predicted value of _Yt ;

Then the QR-CNN-BiLSTM model can be expressed as Estimator of model parameter Ω(τ) at each quantile By minimizing the loss function acquired;

Step 6.4: Model parameter setting and model training;

The network structure of the QR-CNN-BiLSTM model consists of a layer of CNN, a layer of maximum pooling, three layers of BiLSTM, and a layer of fully connected layer.

Set the number of convolution layers to 1, the number of convolution kernels to 64, the convolution kernel size to 4, the convolution boundary processing method to 'same', the activation function to , and the pooling window size to 3;

Set the number of BiLSTM network layers to 3, the number of neurons to 128, and the dropout parameter to 0.2; set the starting point of the quantile to 0.05, the step size to 0.05, and the end point to 1, so the number of neurons in the fully connected layer is 19;

Set the initial learning rate to 0.01, the learning rate decay to 1.5, the minimum learning rate to 10 ^-4 , the maximum number of iterations to 100, the batch parameter to 32, and the optimizer to Adam;

The training set of the similar day data set is input into the constructed QR-CNN-BiLSTM model for model training.

The beneficial effects of the present invention are:

The present invention is based on a photovoltaic power interval probability prediction method based on a deep learning fusion model, and proposes a correlation analysis method based on the kendall rank correlation coefficient, which can determine the model input variables and reduce the invalid information in the historical data to improve the training efficiency; the meteorological factor variables with high correlation with the photovoltaic power are selected as clustering variables, and 7 statistical features such as the average value of the clustering variables are selected as clustering features to comprehensively reflect the fluctuation law and characteristics of the historical data of each day, so that the FCM algorithm can perform efficient clustering, and the clustering algorithm is efficient and reasonable; the QR-CNN-BiLSTM model integrates two deep learning models, CNN and BiLSTM, and has higher prediction accuracy than the traditional single deep learning prediction model. When predicting the future photovoltaic power at a fine time granularity of 5 minutes, the photovoltaic power at an interval of 5 minutes in non-sunny weather presents a more rapid change feature, which can better track the future photovoltaic power trend, and on the basis of meeting the credibility requirements, realizes a high-accuracy measurement of the uncertainty of photovoltaic power prediction, and generates a photovoltaic power prediction interval at a corresponding confidence level, which has practical application value.

BRIEF DESCRIPTION OF THE DRAWINGS

FIG1 is a flow chart of a photovoltaic power interval probability prediction method based on a deep learning fusion model of the present invention;

FIG2 is a flow chart of a fuzzy C-means clustering algorithm in a photovoltaic power interval probability prediction method based on a deep learning fusion model of the present invention;

FIG3 is a diagram showing interval prediction results using the QR-CNN-BiLSTM model in the method of the present invention on a sunny day;

Figure 4 is a graph showing the interval prediction results using the QR-LSTM model on sunny days;

Figure 5 is a graph showing the interval prediction results using the QR-BiLSTM model on sunny days;

FIG6 is a diagram showing interval prediction results when the weather changes from sunny to cloudy using the QR-CNN-BiLSTM model in the method of the present invention;

Figure 7 is a graph showing the interval prediction results when the weather changes from sunny to cloudy using the QR-LSTM model;

Figure 8 is a graph showing the interval prediction results of the QR-BiLSTM model when the weather changes from sunny to cloudy;

FIG9 is a diagram showing interval prediction results using the QR-CNN-BiLSTM model in the method of the present invention on rainy days;

Figure 10 is a graph showing the interval prediction results using the QR-LSTM model on rainy days;

Figure 11 is a graph showing the interval prediction results using the QR-BiLSTM model on rainy days;

FIG12 is a graph showing the probability prediction results of the QR-CNN-BiLSTM model combined with the kernel density estimation method on a sunny day according to the method of the present invention;

FIG13 is a graph showing the probability prediction results of the QR-CNN-BiLSTM model combined with the kernel density estimation method when the weather changes from sunny to cloudy;

FIG14 is a graph showing the probability prediction results of the QR-CNN-BiLSTM model of the method of the present invention combined with the kernel density estimation method on rainy days.

Detailed ways

The present invention is described in detail below with reference to the accompanying drawings and specific embodiments.

The photovoltaic power interval probability prediction method based on the deep learning fusion model of the present invention is shown in FIG1 and is specifically implemented according to the following steps:

Step 1: Obtain pre-processed meteorological element data and historical photovoltaic power data, perform variable correlation analysis on photovoltaic power and meteorological factors, and determine the input variables of the prediction model; specifically:

Step 1.1, select the pre-processed meteorological element data and photovoltaic power generation data, and keep the time resolution of meteorological element variables and photovoltaic power variables consistent;

The time resolution of meteorological element variables and photovoltaic power variables is 5 minutes. Wind speed, relative humidity, ambient temperature, total horizontal radiation, diffuse horizontal radiation, rainfall, wind direction, total inclined radiation and diffuse inclined radiation are selected as the original meteorological element data variables. Since the photovoltaic array does not generate electricity at night, the data with zero power at night are removed, and the data from 6:00 to 19:30 every day are retained as example analysis data.

Find the missing values of historical meteorological data and photovoltaic power generation data, and fill them with interpolation method; use box plot to find the outliers of original historical meteorological data and photovoltaic power generation data, and replace the outliers of data with the upper and lower boundaries of the box plot;

Step 1.2, the Kendall rank correlation coefficient R is used to measure the degree of correlation between multiple meteorological element variables;

The definition of Kendall rank correlation coefficient R is as follows:

Where: P represents the number of consistent pairs; Q represents the number of inconsistent pairs, Represents the total number of observation logs. When two pairs of observations A _i , B _i and A _j , B _j of variables A and B satisfy A _i <B _i and A _j <B _j , the two pairs of observations are said to be consistent or harmonious, otherwise they are inconsistent or disharmonious;

The closer the Kendall rank correlation coefficient R is to 1, the higher the correlation between the meteorological variable and the output power. A positive number indicates a positive correlation, while a negative number indicates a negative correlation.

Step 1.3, select meteorological factors whose absolute value of the kendall rank correlation coefficient R of photovoltaic power is not less than 0.5, and use them as input of the prediction model;

In this example, the kendall rank correlation coefficient R between the photovoltaic power variable and each meteorological element variable is shown in Table 1:

Table 1 Meteorological factors variables

Select multiple meteorological factors with high correlation with photovoltaic power, that is, meteorological factors with an absolute value of the kendall rank correlation coefficient with photovoltaic power of not less than 0.5, and use them as inputs of the prediction model. Therefore, total horizontal radiation, diffuse horizontal radiation, total inclined radiation, and diffuse inclined radiation are selected as meteorological element variables of this embodiment;

Step 2: Select clustering variables and construct statistical characteristics of clustering variables, specifically:

Step 2.1, select the meteorological factor variable with the highest kendall rank correlation coefficient value of photovoltaic power as the clustering variable; in this embodiment, the total horizontal radiation is selected as the clustering variable;

Step 2.2, select the mean, standard deviation, maximum value, number of peaks and valleys, coefficient of variation, kurtosis and skewness of the clustering variable as statistical features;

The definitions of coefficient of variation C, kurtosis K and skewness D are shown in equations (2)-(4) respectively:

In the formula, σ represents the standard deviation of the variable, represents the mean value of the variable, _Xi represents a sample of the variable, and M represents the total number of samples of the variable;

Step 3: Based on the clustering variables and their statistical characteristics selected in step 2, as shown in Figure 2, the fuzzy C-means (FCM) clustering algorithm is used to cluster similar days of photovoltaic historical data to obtain a photovoltaic similar day data set; specifically:

Step 3.1, according to the statistical characteristics of the clustering variables constructed in step 2, calculate the values of the seven statistical characteristics of the clustering variables on each day, and use them as clustering data for clustering using the fuzzy C-means (FCM) clustering algorithm;

Step 3.2, determine the number of data clustering categories c, initialize the cluster center V _i , give the fuzzification parameter m, initialize the membership matrix U ⁽⁰⁾ , and give the termination criterion ε of the algorithm;

Step 3.3: Calculate all cluster centers of the tth iteration according to formula (5) to obtain the cluster center matrix:

Where: _uij is the membership of the i-th sample to the j-th class; _xi is the sample point; m is the membership factor; n is the number of samples; t is the number of iterations; c represents the number of clusters.

Step 3.4: Update the membership matrix U ^(t) . The calculation method is shown in formula (6):

Where: _uij is the membership of the ith sample to the jth class; _xi is the sample point; m is the membership factor; n is the number of samples; t is the number of iterations; c represents the number of clusters; _dij is the distance from the ith sample to the center of the jth cluster.

Step 3.5, calculate ||U ^(t) -U ^(t-1) ||, and verify whether the iteration stop condition ||U ^(t) -U ^(t-1) || < ε is met. If the condition is met, stop the iteration. Otherwise, continue to repeat steps 3.3 and 3.4 until the condition is met. Finally, a similar daily data set under various weather conditions is obtained.

Step 4: Use the min-max normalization method to normalize the photovoltaic similar day data set to eliminate the impact of different variable dimension differences on the prediction results;

The min-max normalization method is shown in formula (7):

In the formula, x ^* represents the data to be standardized, x represents the data after standardization, _xmax represents the maximum value of a variable data, and _xmin represents the minimum value of a variable data.

Step 5: Divide the similar day data sets under various weather conditions into training sets and test sets;

Step 6: Construct a quantile regression (QR)-convolutional neural network (CNN)-bidirectional long short-term memory (BiLSTM) neural network, namely a QR-CNN-BiLSTM interval prediction model, set the parameters of the CNN and BiLSTM models, and set the relevant parameters for model training; input the training set of the similar day data set into the constructed QR-CNN-BiLSTM model for model training; specifically:

Step 6.1: construct a feature map of the normalized similar day data set in the form of a sliding window and input it into the CNN network, and use its convolutional layer and pooling layer to extract the feature vector that characterizes the dynamic change of photovoltaic power;

The neuron output after the convolutional neural network extracts local features is shown in formula (8);

Where: O represents the local output of the neuron; I represents the input of the neuron; l, m, and n represent the three dimensions of the output matrix; i, j, and n represent the length, width, and depth of the convolution kernel K; b _n represents the threshold of the convolution kernel; Represents a matrix multiplication operation.

Step 6.2: Convert the output feature vector into a time series and input it into the BiLSTM network to further capture the long-term dependencies in the time series.

The calculation process of the BiLSTM network is as follows: the forget gate determines which input information will be deleted from the memory cell state, as shown in formula (9);

f _t =σ(W _f ·[h _t-1 ,X _t ]+b _f ) (9)

The output value of the previous moment and the input value of the current moment are input into the input gate, and the output value of the input gate is obtained after calculation, as shown in formula (10);

i _t =σ(W _i ·[h _t-1 ,X _t ])+b _i (10)

The output value at the previous moment and the input value at the current moment are input into the input gate, and the candidate cell state is obtained after calculation, as shown in formula (11);

Update the current cell state, as shown in formula (12);

The output value of the previous moment and the input value of the current moment are input to the output gate, and the output value of the output gate is obtained after calculation, as shown in formula (13);

o _t =σ(W _o ·[h _t-1 ,X _t ]+b _o ) (13)

The output of the output gate is calculated with the cell state to obtain the output value, as shown in formula (14);

h _t = o _t *tanh(C _t ) (14)

Where: f _t is the output of the forget gate at time t; σ and tanh functions are both activation functions; h _t-1 is the data output information at time t-1; X _t is the data input information at time t; W _f , _Wi , W _C , and W _o are weight coefficients; b _f , _bi , b _C , and b _o are bias parameters; i _t and represents the output of the input at time t; C _t-1 is the cell state at time t-1; C _t is the cell state at time _t ; h _t is the data output information at time t; o _t represents the output after activation by the activation function Sigmoid at time t.

Input the forward data into the forward LSTM layer to get the output of the forward LSTM layer. Then input the data into the reverse LSTM layer in reverse direction, get the reverse output, and then reverse the output again to get the output of the reverse LSTM layer. Finally, linearly superimpose the output of the forward LSTM layer and the output of the reverse LSTM layer according to certain weights to get the output result.

Step 6.3: Introduce the QR model and combine it with CNN-BiLSTM. The QR model and CNN-BiLSTM model are fused in the form of loss function.

set up represents the CNN-BiLSTM point prediction model, where _Xt is the model input, i.e., the independent variable, Ω is the model parameter of CNN-BiLSTM, and _Yt is the dependent variable. is the predicted value of _Yt .

Then the QR-CNN-BiLSTM model can be expressed as Estimator of model parameter Ω(τ) at each quantile By minimizing the loss function acquired;

Step 6.4: Model parameter setting and model training.

The network structure of the QR-CNN-BiLSTM model consists of a layer of CNN, a layer of maximum pooling, three layers of BiLSTM, and a layer of fully connected layer.

Set the number of convolution layers to 1, the number of convolution kernels to 64, the convolution kernel size to 4, the convolution boundary processing method to 'same', the activation function to , and the pooling window size to 3;

Set the number of BiLSTM network layers to 3, the number of neurons to 128, and the dropout parameter to 0.2; set the starting point of the quantile to 0.05, the step size to 0.05, and the end point to 1, so the number of neurons in the fully connected layer is 19.

Set the initial learning rate to 0.01, the learning rate decay to 1.5, the minimum learning rate to 10 ^-4 , the maximum number of iterations to 100, the batch parameter to 32, and the optimizer to Adam;

Input the training set of similar day data set into the constructed QR-CNN-BiLSTM model for model training;

Step 7: Input the test set data into the QR-CNN-BiLSTM model trained in step 6 to predict the photovoltaic power range;

Step 8: Denormalize the interval prediction results to make them physically meaningful;

Step 9: Based on the denormalized photovoltaic power interval prediction result data obtained in step 8, a kernel density estimation algorithm optimized by cross-validation and grid search methods is used to generate photovoltaic probability prediction results on the test set;

Specifically, for a specific future photovoltaic power prediction point, the QR-CNN-BiLSTM fusion model in step 7 is applied to obtain a set of N samples under N conditional quantiles, that is, the vector containing N samples is Its probability density function can be obtained by the kernel density estimation method. The KDE calculation of this vector is shown in formula (15):

Where: N is the total number of samples; B is the optimal bandwidth determined by the cross-validation grid search method, and B>0; K is a kernel function. Gaussian kernel is the kernel function K used in the present invention, and the Gaussian kernel function is shown in formula (16):

The cross-validation and grid search methods proposed in the present invention optimize kernel density estimation, solve the problem of difficult bandwidth selection in kernel density estimation, and can generate high-quality probability prediction results.

The photovoltaic power point prediction, interval prediction and probability prediction results are evaluated. The root mean square error e _RMSE and mean absolute percentage error e _MAPE are used to evaluate the point prediction results; the interval comprehensive evaluation index I _WC is used to evaluate the interval prediction results; the continuous graded probability score P _CRPS is used to evaluate the probability prediction results; as shown in the following formula:

I _WC = I _PINAW /I _PICP

in:

Where: _Pri represents the power observation value, _Ppi represents the power prediction value, and N is the total number of predicted future photovoltaic power points. _Sn is a logical value. When the observed value falls within the prediction interval, _Sn takes 1, otherwise _Sn takes 0; E is the difference between the maximum and minimum values of the observed value; _Pupi and _Pdowni are the upper and lower bounds of the prediction interval respectively; p(x) represents the probability density function; F( _Ppi ) represents the cumulative distribution function of _Ppi ; H( _Ppi - _Pri ) is the step function.

The photovoltaic power interval probability prediction method based on the deep learning fusion model of the present invention first uses the correlation coefficient method to screen meteorological factors to reduce the model prediction error caused by too many irrelevant features; then selects highly correlated meteorological factor variables as clustering variables and constructs their statistical characteristics, and uses the FCM clustering algorithm to cluster to obtain similar day data sets to lay the foundation for further improving the prediction accuracy; the QR-CNN-BiLSTM model integrates two deep learning models, CNN and BiLSTM, and has higher prediction accuracy than the traditional single deep learning prediction model, and can generate higher quality interval prediction results; cross-validation and grid search methods optimize kernel density estimation, solve the problem of difficult bandwidth selection in kernel density estimation, and this method can generate high-quality probability prediction results.

Example

The photovoltaic power data of the photovoltaic array produced by a manufacturer at the Alice Springs site of the Australian Desert Knowledge Solar Center (DKASC) from 2019 to 2020 and the four meteorological factor data obtained by the above screening are used as simulation data. The time resolution of this data set is 5 minutes. Since the photovoltaic array does not generate electricity at night, the data with zero power at night are removed, and the data from 6:00 to 19:30 every day are retained as the case analysis data.

The fuzzy C-means clustering algorithm is used to finally divide the data set into three categories: sunny, sunny to cloudy, and rainy similar day data sets. Each similar day data set is divided into a training set and a test set, with the training set accounting for 0.7. The test set selects the 30-day similar day data that is close in time, and the training set selects the 70-day similar day data that is closest to the test set in the similar day data set.

In order to illustrate the advantages of the proposed model in the short-term interval and probability prediction of photovoltaic power, the prediction effects of the proposed QR-CNN-BiLSTM model, QR-LSTM and QR-BiLSTM models were compared under three weather types. One day was randomly selected in each weather type for visualization analysis, as shown in Figure 3-11, which shows the prediction results of each model point and 95% confidence level interval under sunny, sunny to cloudy and rainy weather types. Figures 12-14 show the probability prediction results of the QR-CNN-BiLSTM model combined with kernel density estimation under sunny, sunny to cloudy and rainy weather types (the probability prediction results of 9 points were equally spaced from 164 points for display). The comparison of evaluation indicators is shown in Tables 2, 3 and 4.

Table 2 Evaluation indicators of the model

Table 3 Evaluation indicators of the model

Table 4 Evaluation indicators of the model

From Table 2, it can be seen that on sunny days, the coverage of the prediction intervals of each model is close to 100%. The prediction interval width of the QR-CNN-BiLSTM model is significantly lower than that of other models. The PINAW index is 72.77% lower than that of QR-LSTM and 66.26% lower than that of QR-BiLSTM. At the same time, the interval comprehensive evaluation index value I _WC of the QR-CNN-BiLSTM model is also the smallest. The I _WC index is 72.62% lower than that of QR-LSTM and 66.08% lower than that of QR-BiLSTM. Therefore, its interval prediction performance is the best. The probability evaluation index CRPS value of QR-CNN-BiLSTM is the smallest, so its probability prediction performance is also the best. From the deterministic evaluation index, it can be seen that the point prediction performance of the QR-CNN-BiLSTM model is also better.

From Table 3, it can be seen that when the weather changes from sunny to cloudy, under the premise that the PICP values of the prediction intervals of each model are all >95%, the PINAW of the prediction interval of the QR-CNN-BiLSTM model is significantly reduced, which is 28.34% lower than that of QR-LSTM and 25.62% lower than that of QR-BiLSTM; at the same time, the I _WC value of the interval comprehensive evaluation index of the QR-CNN-BiLSTM _model is also the smallest, which is 28.80% lower than that of QR-LSTM and 26.09% lower than that of QR-BiLSTM, so its interval prediction performance is the best. The CRPS value of the probability evaluation index of QR-CNN-BiLSTM is the smallest, so its probability prediction performance is also the best. From the deterministic evaluation index, it can be seen that the point prediction performance of the QR-CNN-BiLSTM model is also better.

It can be seen from Table 4 that on rainy days, under the premise that the PICP values of the prediction intervals of each model are all >95%, the PINAW of the prediction interval of the QR-CNN-BiLSTM model is significantly reduced, which is 7.98% lower than that of QR-LSTM and 4.47% lower than that of QR-BiLSTM. At the same time, the I _WC value of the interval comprehensive evaluation index is also the smallest. The I _WC index is 6.81% lower than that of QR-LSTM and 4.08% lower than that of QR-BiLSTM, which significantly reduces the uncertainty of the photovoltaic power interval prediction. Therefore, the QR-CNN-BiLSTM interval prediction performance is the best. The CRPS value of the QR-CNN-BiLSTM probability evaluation index is the smallest, so its probability prediction performance is also the best. From the deterministic evaluation index, it can be seen that the QR-CNN-BiLSTM model has better point prediction performance. In summary, the point prediction, interval prediction and probability prediction performance of the QR-CNN-BiLSTM model are all superior.

Claims (4)

1. A photovoltaic power interval probability prediction method based on a deep learning fusion model is characterized by being implemented in the following steps: Step 1: Obtain pre-processed meteorological element data and historical photovoltaic power data, perform variable correlation analysis on photovoltaic power and meteorological factors, and determine the input variables of the prediction model; Step 2: Select clustering variables and construct statistical characteristics of clustering variables. Step 3: Based on the clustering variables and their statistical characteristics selected in step 2, the fuzzy C-means clustering algorithm is used to cluster similar days of photovoltaic historical data to obtain a photovoltaic similar day data set; Step 4: Use the min-max normalization method to normalize the photovoltaic similar day data set; Step 5: Divide the similar day data sets under various weather conditions into training sets and test sets; Step 6: Build the QR-CNN-BiLSTM interval prediction model, set the parameters of the CNN and BiLSTM models, and set the relevant parameters for model training; input the training set of the similar day data set into the model for training; specifically: Step 6.1: construct a feature map of the normalized similar day data set in the form of a sliding window and input it into the CNN network, and use its convolutional layer and pooling layer to extract the feature vector that characterizes the dynamic change of photovoltaic power; Step 6.2: Convert the output feature vector into a time series and input it into the BiLSTM network; Step 6.3: Introduce the QR model and combine it with CNN-BiLSTM. The QR model and CNN-BiLSTM model are fused in the form of loss function. set up represents the CNN-BiLSTM point prediction model, where _Xt is the model input, i.e., the independent variable, Ω is the model parameter of CNN-BiLSTM, and _Yt is the dependent variable. is the predicted value of _Yt ; Then the QR-CNN-BiLSTM model can be expressed as Estimator of model parameter Ω(τ) at each quantile By minimizing the loss function acquired; Step 6.4: Model parameter setting and model training; The network structure of the QR-CNN-BiLSTM model consists of a layer of CNN, a layer of maximum pooling, three layers of BiLSTM, and a layer of fully connected layer. Set the number of convolution layers to 1, the number of convolution kernels to 64, the convolution kernel size to 4, the convolution boundary processing method to 'same', the activation function to , and the pooling window size to 3; Set the number of BiLSTM network layers to 3, the number of neurons to 128, and the dropout parameter to 0.2; set the starting point of the quantile to 0.05, the step size to 0.05, and the end point to 1, so the number of neurons in the fully connected layer is 19; Set the initial learning rate to 0.01, the learning rate decay to 1.5, the minimum learning rate to 10 ^-4 , the maximum number of iterations to 100, the batch parameter to 32, and the optimizer to Adam; Input the training set of similar day data set into the constructed QR-CNN-BiLSTM model for model training; Step 7: Input the test set data into the QR-CNN-BiLSTM interval prediction model trained in step 6 to perform photovoltaic power interval prediction; Step 8: Denormalize the interval prediction results to make them physically meaningful; Step 9: Use the kernel density estimation algorithm optimized by cross-validation and grid search methods to generate photovoltaic probability prediction results on the test set. 2. The photovoltaic power interval probability prediction method based on deep learning fusion model according to claim 1 is characterized in that in step 1, specifically: Step 1.1, select the pre-processed meteorological element data and photovoltaic power generation data; The time resolution of meteorological element variables and photovoltaic power variables is 5 min, and wind speed, relative humidity, ambient temperature, total horizontal radiation, diffuse horizontal radiation, rainfall, wind direction, total oblique radiation and diffuse oblique radiation are selected as the original meteorological element data variables; Step 1.2, the Kendall rank correlation coefficient R is used to measure the degree of correlation between multiple meteorological element variables; Step 1.3: Select meteorological factors whose absolute value of the Kendall rank correlation coefficient R of photovoltaic power is not less than 0.5 and use them as input to the prediction model. 3. The photovoltaic power interval probability prediction method based on deep learning fusion model according to claim 2 is characterized in that in step 2, specifically: Step 2.1, select the meteorological factor variable with the highest kendall rank correlation coefficient value of photovoltaic power as the clustering variable; Step 2.2: Select the mean, standard deviation, maximum value, number of peaks and troughs, coefficient of variation, kurtosis and skewness of the clustering variable as statistical features. 4. The photovoltaic power interval probability prediction method based on deep learning fusion model according to claim 1 is characterized in that in step 3, specifically: Step 3.1, based on the statistical characteristics of the clustering variables constructed in step 2, calculate the values of the seven statistical characteristics of the clustering variables on each day; Step 3.2, determine the number of data clustering categories c, initialize the cluster center V _i , give the fuzzification parameter m, initialize the membership matrix U ⁽⁰⁾ , and give the termination criterion ε of the algorithm; Step 3.3: Calculate all cluster centers of the tth iteration according to formula (5) to obtain the cluster center matrix: Where: _uij is the membership of the i-th sample to the j-th class; _xi is the sample point; m is the membership factor; n is the number of samples; t is the number of iterations; c represents the number of clusters; Step 3.4: Update the membership matrix U ^(t) . The calculation method is shown in formula (6): Where: u _ij is the membership of the ith sample to the jth class; _xi is the sample point; m is the membership factor; n is the number of samples; t is the number of iterations; c is the number of clusters; d _ij is the distance from the ith sample to the center of the jth cluster; Step 3.5, calculate ||U ^(t) -U ^(t-1) ||, and verify whether the iteration stop condition ||U ^(t) -U ^(t-1) || < ε is met. If the condition is met, stop the iteration; otherwise, continue to repeat steps 3.3 and 3.4 until the condition is met, and finally obtain a similar daily data set under various weather conditions.