Short-term photovoltaic output interval prediction method, model and storage medium

Abstract

The invention provides a short-term photovoltaic output interval prediction method and model and a storage medium, and the method comprises the steps: constructing different CNN-BiLSTM combination prediction models based on a plurality of modal components and optimal input feature combinations; performing hyper-parameter optimization on the CNN-BiLSTM combined prediction model by using a Bayesian optimization algorithm to obtain an optimal hyper-parameter; and based on the optimal hyper-parameters corresponding to the different components, photovoltaic output short-term prediction is performed on the different components by using a CNN-BiLSTM combined prediction model to obtain photovoltaic output predicted values of the different components. According to the invention, a Bootstrap self-service method is adopted to reconstruct an error sample set of photovoltaic output point prediction, so that the prediction accuracy of the photovoltaic output point can be improved on the premise of not needing any hypothesis. The interval prediction of the photovoltaic output is realized, and the method has a good prediction effect especially for time series data with relatively large volatility, such as the photovoltaic output.

Description

Technical Field

[0001] The present application belongs to the technical field of power systems, and in particular, relates to a short-term photovoltaic output interval prediction method, model and storage medium. Background Art

[0002] With the steady progress of the construction of new power systems, its "double high" characteristics of "high proportion of power electronic equipment access" and "high proportion of renewable energy penetration" are becoming more and more obvious. Photovoltaic power generation, as one of the main forms of renewable energy utilization, has the advantages of wide distribution, clean and environmental protection, and its position in new energy power generation is becoming more and more important. However, photovoltaic output is affected by many factors such as irradiance, ambient temperature, and photovoltaic module conversion efficiency. It has the characteristics of strong randomness, large volatility, and obvious intermittency. In addition, when using deep learning algorithms, especially hybrid neural network algorithms in deep learning, to predict photovoltaic output, it is difficult to obtain its appropriate high-dimensional hyperparameters, which will greatly affect the prediction performance of the model. Moreover, the existing deterministic prediction methods cannot provide the best guidance for power system scheduling decisions or risk assessment.

[0003] With a high proportion of photovoltaic power generation connected to the power system, the dispatch planning and safe and stable operation of the power grid are facing huge challenges. Accurate photovoltaic output short-term interval prediction technology can effectively alleviate the impact of photovoltaic output uncertainty on the power grid, and help build a comprehensive security defense system for the power system and carry out power system risk assessment and other work. Summary of the invention

[0004] In view of the limitations of existing methods, a short-term photovoltaic output interval prediction method, model and storage medium are proposed. The complex signal is decomposed into multiple modal components by ICEEMDAN, making each component easier to analyze and predict. The maximum mutual information coefficient is used for feature screening to ensure that the features of the model input are highly correlated. The two deep learning models of CNN and BiLSTM are combined to give full play to the ability to capture spatial and temporal features. This application also uses the Bayesian optimization algorithm for hyperparameter optimization to improve the prediction accuracy of the model. The prediction error is reconstructed by the Bootstrap algorithm to obtain a more reliable prediction interval.

[0005] To achieve the above objectives, the present application provides a short-term photovoltaic output interval prediction method, model and storage medium, specifically including:

[0006] S1: Obtain the photovoltaic output curve data of the photovoltaic power station within time T, pre-process the original data, and obtain a real data set;

[0007] S2: performing modal decomposition on the real data set based on an improved adaptive noise complete set empirical mode decomposition algorithm to obtain multiple modal components;

[0008] S3: using the maximum mutual information coefficient method to perform feature screening on the multiple modal components respectively, and screen out optimal input feature data for different modal components;

[0009] S4: constructing different CNN-BiLSTM combination prediction models based on the multiple modal components and their optimal input feature combinations;

[0010] S5: Using the Bayesian optimization algorithm to optimize the hyperparameters of the CNN-BiLSTM combined prediction model to obtain the optimal hyperparameters;

[0011] S6: Based on the optimal hyperparameters corresponding to different components, the CNN-BiLSTM combined prediction model is used to perform short-term photovoltaic output prediction for different components to obtain photovoltaic output prediction values ​​for different components;

[0012] S7: Accumulating the photovoltaic output prediction values ​​of the different components to obtain a point prediction result of the final daily photovoltaic output, and reconstructing its prediction error based on the Bootstrap algorithm to obtain a final photovoltaic output prediction interval.

[0013] Preferably, the step S2 specifically includes:

[0014] S21: Add a set of white noise ω to the original signal x(t) i (t), construct the sequence x i (t), and obtain the first group of residual components δ1(t):

[0015] x i (t) = x(t) + γ0E1(ω i (t));

[0016] δ1(t)=<M(x1(t))> ;

[0017] Where x(t) represents the original signal to be decomposed; ω i (t) indicates the i-th time Gaussian white noise is added; x i (t) is the new signal constructed after adding white noise for the ith time; γ k-1 is the adaptive coefficient for solving the k-th order modal component; E k (·) represents the kth-order modal component generated by EMD decomposition; δ k represents the k-th order residual component obtained by ICEEMDAN; M(·) represents the local mean of the signal; <·> represents the overall mean;

[0018] S22: Calculate the first-order modal component f1:

[0019] f1=x(t)-δ1;

[0020] S23: Continue to add white noise and express the second set of residual components δ2(t) as δ1+γ1E2ω i (t) mean, and obtain the second-order modal component f2:

[0021] f2=δ1-δ2=δ1- <M(δ1+γ1E2ω i (t))>;

[0022] S24: Repeat step S23 to obtain the k-th order residual component δ k With the modal component f k :

[0023] δ k = <M(δ k-1 +γ k-1 E k ω i (t))>;

[0024] f k =δ k-1 -δ k =δ k-1 - <M(δ k-1 +γ k-1 E k ω i (t))>;

[0025] S25: until all modal components and residual components are obtained, the modal decomposition is completed

[0026] Preferably, the step S3 specifically includes:

[0027] S31: For a binary data set C(x,y), divide it into a and b intervals in the X direction and Y direction respectively, then a×b grids G are formed in the two-dimensional plane. Let the set of grids G formed by different division methods be Ω, C|G represents the distribution of the data set C on the network G, then:

[0028]

[0029] In the formula, I * (C,a,b) is the maximum mutual information of X and Y under data set C;

[0030] S32: The feature matrix composed of the maximum normalized mutual information of the data set C is represented as M, then the element m ab Defined as:

[0031]

[0032] S33: The final MIC value is obtained as follows:

[0033]

[0034] Where N is the total number of samples; C(N) is the sample number function, generally C(N) = N 0.6 ;

[0035] S34: MIC correlation values ​​between different features and different modal components are calculated according to S31-S33, a threshold is set, features with MIC correlation values ​​less than the threshold are eliminated, and the remaining features are used as the optimal input feature data corresponding to different modal components.

[0036] Preferably, the step S4 specifically includes:

[0037] S41: Dividing the optimal input feature data of the different modal components obtained in step S2 and the different components obtained in step S3 into a training set and a test set according to a certain ratio;

[0038] S42: Constructing convolutional neural network layers in sequence: convolutional layer, pooling layer, convolutional layer, and pooling layer. The training set is constructed as a feature map in the form of a sliding window and input into the convolutional neural network layer. The hidden layer is used to extract feature vectors that can characterize photovoltaic output information.

[0039] S43: construct a BiLSTM network layer, which is composed of forward and reverse LSTM neural networks; the feature vector extracted by the convolutional neural network layer will be input into the BiLSTM neural network after being flattened by the Flatten layer, and the BiLSTM captures the bidirectional time series correlation of the photovoltaic output data; wherein, the specific expression of the BiLSTM neural network is as follows:

[0040]

[0041] Where: x t , A t , B t and t represents the input data at time t, the hidden layer of the forward iteration, the hidden layer of the backward iteration, and the output data. w1, w2, …, w6 are the weights of the corresponding layers, and η1, η2, η3 are the activation functions of the corresponding layers.

[0042] Preferably, the step S5 specifically includes:

[0043] S51: Determine hyperparameter space: Define the hyperparameter space of the CNN-BiLSTM model;

[0044] S52: Select Gaussian process as the proxy model: initialize the Gaussian process proxy model as the probability model of the objective function;

[0045] S53: Adopting Gaussian process posterior maximization strategy as the selection sampling strategy to select the next sample point in the hyperparameter space to evaluate the objective function;

[0046] S54: Randomly initialize the proxy model and the sampling strategy;

[0047] S55: Iterative optimization to obtain the optimal hyperparameter combination

[0048] Wherein, the step S55 specifically includes:

[0049] S551: Based on the posterior probability maximization strategy of the Gaussian process surrogate model, select the next hyperparameter configuration to maximize the potential performance improvement;

[0050] S552: Evaluate the configuration on the real model to obtain a performance indicator, i.e., an error in photovoltaic output prediction;

[0051] S553: ​​Use the evaluation results to update the Gaussian process surrogate model and update the estimates of the mean and variance;

[0052] S554: reselect the next configuration based on the posterior probability estimate;

[0053] S555: When a predetermined number of iterations is reached or the performance converges, the hyperparameter configuration with the maximum estimated value in the proxy model is selected as the optimal hyperparameter combination finally found.

[0054] Preferably, the step S7 specifically includes:

[0055] S71: Accumulating the photovoltaic output prediction values ​​of different components in step S6 to obtain the final photovoltaic output point prediction result;

[0056] S72: subtracting the actual photovoltaic output data from the photovoltaic output point prediction result to obtain a final photovoltaic output point prediction error data set;

[0057] S73: resampling to create multiple new data sets; wherein the size of the data sets is the same as the original data;

[0058] S74: Calculate statistics, for each new data set in S73, calculate its statistics;

[0059] S75: repeating steps S73 and S74 to form a series of different statistical values ​​and construct a distribution of the statistics;

[0060] S76: Using the distribution of the statistic calculated in S75 and according to the set confidence level, a final photovoltaic output prediction interval is calculated.

[0061] To achieve the above objectives, the present application also provides a short-term photovoltaic output interval prediction model, which is used to implement the short-term photovoltaic output interval prediction method as described above.

[0062] To achieve the above objectives, the present application also provides a computer-readable storage medium, which stores computer-executable instructions. When the computer-executable instructions are executed by a control processor, a short-term photovoltaic output interval prediction method as described above is implemented.

[0063] The present invention uses the ICEEMDAN time-frequency analysis method to perform modal decomposition and stabilization processing on the original photovoltaic output data to obtain multiple modal components. For each modal component obtained by decomposition, the MIC method is used to analyze the correlation between each component and the multidimensional input feature, and the input features with high correlation are respectively screened out as the input features of different modal components. Then, the CNN-BiLSTM combined prediction model optimized by BO is used to realize the high-precision photovoltaic output point prediction of each modal component respectively, and the point prediction results of all modal components are accumulated to obtain the final photovoltaic output point prediction result. Finally, the Bootstrap method is used to reconstruct the point prediction error data set of photovoltaic output, and the interval prediction result of photovoltaic output is obtained, so as to realize the quantification of photovoltaic output uncertainty and provide more effective guidance for power system scheduling.

[0064] The prediction method of the present invention studies the high volatility and strong randomness characteristics of photovoltaic output, and adopts the ICEEMDAN method to perform modal decomposition of photovoltaic output, thereby effectively reducing the impact of the high volatility and strong randomness of photovoltaic output on the accurate prediction of photovoltaic output.

[0065] The present invention adopts MIC to realize the feature screening work for each modal component, fully considers the influence of meteorological factors on photovoltaic output, and the time series correlation between photovoltaic output data and its historical data, so as to further improve the prediction accuracy of photovoltaic output.

[0066] The present invention adopts the combined prediction model CNN-BiLSTM to realize the point prediction of photovoltaic output, combining the powerful feature extraction ability of CNN and the excellent fitting ability of BiLSTM for the bidirectional temporal correlation of time series data. In addition, BO is used to realize the optimization of high-dimensional hyperparameters of the CNN-BiLSTM algorithm, thereby exerting the best prediction performance of the CNN-BiLSTM combined prediction model.

[0067] The present invention adopts the Bootstrap method to reconstruct the error sample set of photovoltaic output point prediction, which can realize the interval prediction of photovoltaic output without any assumptions, especially for time series data with large volatility such as photovoltaic output, and has a good prediction effect. BRIEF DESCRIPTION OF THE DRAWINGS

[0068] Figure 1 It is a short-term interval prediction flow chart of photovoltaic output;

[0069] Figure 2 This is a schematic diagram of the LSTM network structure;

[0070] Figure 3 This is a schematic diagram of the BiLSTM network structure;

[0071] Figure 4 is a schematic diagram of prediction results of an embodiment point;

[0072] Figure 5 is a schematic diagram of interval prediction results of an embodiment; DETAILED DESCRIPTION

[0073] In order to make the purpose, technical scheme and advantages of the embodiments of the present application clearer, the technical scheme will be clearly and completely described below in conjunction with the embodiments of the present application. Obviously, the described embodiments are part of the embodiments of the present application, rather than all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of this application.

[0074] Embodiment 1:

[0075] As attached Figure 1 As shown, in order to solve the above technical problems, the present application provides a short-term photovoltaic output interval prediction method, which mainly includes the following steps:

[0076] 1) Based on the photovoltaic output curve data of a photovoltaic power station within time T, Lagrange interpolation is performed on the missing data in the original data, and bad data is automatically screened and eliminated. Finally, the data of the photovoltaic output zero value period at night are eliminated, and only the data with photovoltaic power during the day are retained.

[0077] 2) Use the ICEEMDAN time-frequency analysis algorithm to perform modal decomposition on photovoltaic output data. The main steps are as follows:

[0078] 2.1) Determine the ratio of the standard deviation of the additional noise to the standard deviation of the input signal, Nstd, the number of ensemble averages, N, and the maximum number of iterations, M, of the ICEEMDAN algorithm.

[0079] 2.2) Add a set of white noise ω to the original signal x(t) i (t), construct the sequence x i (t), and obtain the first group of residual components δ1(t):

[0080] x i (t) = x(t) + γ0E1(ωi (t)) (1)

[0081] δ1(t)=<M(x1(t))> (2)

[0082] Where x(t) represents the original signal to be decomposed; ω i (t) indicates the i-th time Gaussian white noise is added; x i (t) is the new signal constructed after adding white noise for the ith time; γ k-1 is the adaptive coefficient for solving the k-th order modal component; E k (·) represents the kth-order modal component generated by EMD decomposition; δ k represents the k-th order residual component obtained by ICEEMDAN; M(·) represents the local mean of the signal; <·> represents the overall mean;

[0083] 2.3) Calculate the first-order modal component f1

[0084] f1=x(t)-δ1 (3)

[0085] 2.4) Continue to add white noise and express the second set of residual components δ2(t) as δ1+γ1E2ω i (t) mean, and obtain the second-order modal component f2:

[0086] f2=δ1-δ2=δ1- <M(δ1+γ1E2ω i (t))> (4)

[0087] 2.5) Repeat step 2.4) to get the k-th order residual component δ k With the modal component f k :

[0088] δ k = <M(δ k-1 +γ k-1 E k ω i (t))> (5)

[0089] f k =δ k-1 -δ k =δ k-1 - <M(δ k-1 +γ k-1 E k ω i (t))> (6)

[0090] 2.6) until all modal components and residual components are obtained, and the modal decomposition is completed.

[0091] 3) For all the modal components obtained, the maximum mutual information coefficient method (MIC) is used to perform feature screening on them respectively, and the optimal input feature combination is selected for different components. The specific steps are as follows:

[0092] 3.1) For each binary data set C(x,y) consisting of each modal component and each input feature to be screened, divide it into a and b intervals in the X direction and Y direction respectively, then a×b grids G are formed in the two-dimensional plane. Let the set of grids G formed by different division methods be Ω, C|G represents the distribution of the data set C on the network G, then:

[0093]

[0094] In the formula, I * (C,a,b) is the maximum mutual information of x and y in data set C

[0095] 3.2) The feature matrix composed of the maximum normalized mutual information of the data set C is represented as M, then the element m ab Defined as:

[0096]

[0097] 3.3) The final MIC value is:

[0098]

[0099] Where N is the total number of samples; C(N) is the sample number function, generally C(N) = N 0.6 .

[0100] 3.4) According to equations (8) to (10), the MIC correlation values ​​between different features and different modal components are calculated, the threshold is set to 0.6, the features with smaller correlation are screened out, and the remaining features are used as the final input features of the corresponding modal components.

[0101] 4) Based on the input features of different components and photovoltaic output data obtained by screening, different CNN-BiLSTM combined prediction models are constructed. The specific steps are as follows:

[0102] 4.1) Divide the training set and test set:

[0103] The input feature data sets of different modal components obtained in step 2) and different components obtained in step 3) are divided into a training set and a test set in a ratio of 8:2. Considering the temporal correlation of the modal components, the input features to be screened include both the meteorological characteristics provided by NWP and the historical data values ​​of the modal components, such as the modal component values ​​24 hours or 48 hours ago.

[0104] 4.2) Construct the input data into a sliding window form:

[0105] The amount of data for two days after excluding the zero-value data of photovoltaic output at night is used as the length of the sliding window, and the input features plus the total number of modal components to be predicted are used as the sliding window width, so as to construct the input data into a sliding window data set. At the same time, the prediction method for the modal component is a single-step prediction, that is, the sliding window should contain the input features and modal components of the two days after excluding the photovoltaic output at night data, and the data of the first two days in the historical data are put into the window as training data, and the modal component value at the first moment of the third day is used as the label value. When predicting the modal component value at the second moment of the third day, the window also slides for one moment. At this time, the first modal component data in the window will be deleted, and the modal component value at the first moment of the third day will be added to the end of the window. This cycle continues, and all data are input into the model in a window sliding manner.

[0106] 4.2) Construct a convolutional neural network layer for feature extraction:

[0107] The convolutional neural network model consists of three layers: input layer, hidden layer and output layer. The hidden layer is used as a feature extraction layer, which includes two convolution layers and two pooling layers. The construction order is: convolution layer-pooling layer-convolution layer-pooling layer. The training sample set will be constructed as a feature map in the form of a sliding window and input into the CNN network, and the hidden layer of CNN will be used to fully extract the feature vector that can characterize the photovoltaic output information.

[0108] 4.3) Construct BiLSTM network layer for photovoltaic output prediction:

[0109] The BiLSTM network layer is composed of forward and reverse LSTM neural networks. The feature vector extracted by CNN will be input into the BiLSTM neural network after being flattened by the Flatten layer. BiLSTM can fully capture the bidirectional time series correlation of photovoltaic output data. LSTM mainly adds three gating units based on RNN, namely: forget gate, input gate, and output gate. The forget gate will forget the useless information output at time t-1 and retain the useful information. The forgotten part is determined by the forget gate state f at time t. t and the cell state S at time t-1 t-1 The input gate is responsible for reading the input x at time t t and the intermediate state h at time t-1 t-1 , x t With h t-1 The i generated by the sigmoid function and tanh function in the output gate t and c tThe output gate is mainly used to process the data and pass important information to the next moment. t The current unit state S t After being processed by the tanh function, the state o of the output gate t Joint decision.

[0110] 4.3.1) As attached Figure 2 As shown, the specific expression of the LSTM neural network is as follows:

[0111]

[0112] In the formula, x t is the input at time t; h t-1 is the intermediate state output at time t-1; i t is the input gate state at time t; c t is the state of the input node; σ is the sigmoid activation function; tanh represents the hyperbolic tangent activation function; ⊙ is the Hadamard product; W fx and W fh , W ix and W ih , W cx and W ch , W ox and W oh They are respectively the forget gate, input gate, input node, output gate and x t and h t-1 The multiplied weight matrix; b f 、b o 、b i 、b c They are the biases of the forget gate, input gate, input node, and output gate respectively.

[0113] 4.3.2) Further, as shown in the attached Figure 3 As shown, the specific expression of the BiLSTM neural network is as follows:

[0114]

[0115] Where: x t , A t , B t and t represents the input data at time t, the hidden layer of the forward iteration, the hidden layer of the backward iteration, and the output data. w1, w2, …, w6 are the weights of the corresponding layers, and η1, η2, η3 are the activation functions of the corresponding layers.

[0116] 5) Use the Bayesian optimization algorithm (BO) to optimize the hyperparameters of the CNN-BiLSTM constructed by different components to obtain the optimal hyperparameters. The specific steps of using BO to optimize the hyperparameters of CNN-BiLSTM are as follows:

[0117] 5.1) Determine the hyperparameter space:

[0118] Define the hyperparameter space of the CNN-BiLSTM model, that is, the hyperparameters to be optimized. The hyperparameters to be optimized for the CNN layer are: the number of channels of the first convolution kernel C1, the number of channels of the second convolution kernel C2, the size of the first pooling layer P1, and the size of the second pooling layer P2. The hyperparameters to be optimized for BiLSTM are: the number of neurons in the first hidden layer of the BiLSTM layer N1, the number of neurons in the second hidden layer N2, and the optimization range of the BiLSTM learning rate l are all [2,5]. The optimization ranges of C1 and C2 are both [2,5], the optimization ranges of P1 and P2 are both [1,4], the optimization ranges of N1 and N2 are both [16,256], and the optimization range of l is [10 -6 ,0.1]. The number of convolution kernel channels of the first and second layers of CNN are set to 64 and 128 respectively, and MSE Loss is selected as the prediction model loss function of BiLSTM. The optimizer uses Adam optimizer, the number of iterations is set to 300, and the early termination strategy is adopted during the iteration process. If the loss function is not optimized after more than 30 iterations, the hyperparameter optimization is terminated in advance.

[0119] The mathematical expression of the MSE Loss function is as follows:

[0120]

[0121] In the formula, x i 、x i ' are samples of X and X' respectively, and i∈n, (n is the sample set), m is the sequence dimension, that is, x i =[x i1 ,x i2 ...x im ], x i '=[x i '1,x i '2...x i ' m ]

[0122] The mathematical expression of the optimizer Adam model is:

[0123]

[0124] In the formula, Respectively represent the weights and biases of the network parameters at the nth iteration, and are the second-order gradient momentum accumulated by the loss function in the first n iterations, β1 and β2 are the momentum gradient coefficients, α is the learning rate, is the gradient correction, and ε is a small constant to prevent the denominator from being zero.

[0125] 5.2) Choose Gaussian process as the surrogate model:

[0126] Initialize the Gaussian process surrogate model as a probabilistic model of the objective function. This model is used to model the relationship between hyperparameters and performance and estimate uncertainty. The mathematical expression of the Gaussian process surrogate model is as follows:

[0127] f(x)~GP(μ(x),σ(x,x'))(14)

[0128] Where μ(x) is the mean function, which is usually assumed to be zero and represents the expected value of the function under different input values. σ(x, x') is the covariance function, also known as the kernel function, which is used to describe the correlation between different input values.

[0129] 5.3) Select sampling strategy:

[0130] The Gaussian process posterior maximization strategy is used to select the next sample point in the hyperparameter space to evaluate the objective function. The mathematical expression of the Gaussian process posterior distribution is as follows:

[0131] f(x*)|D~N(μ(x*),σ 2 (x*))(15)

[0132] Where μ(x*) is the mean under input f(x*), σ 2 (x*) is the variance under the input f(x*), and D represents the observed data, that is, a set of hyperparameter configurations and the corresponding objective function values.

[0133] 5.4) Randomly initialize the surrogate model and sampling strategy.

[0134] 5.5) Iterate the optimization and repeat the following steps until the stopping condition is reached:

[0135] 5.5.1) According to the posterior probability maximization strategy of the Gaussian process surrogate model, the next hyperparameter configuration is selected to maximize the potential performance improvement;

[0136] 5.5.2) Evaluate the configuration on the real model and obtain the performance index, i.e. the error of PV output prediction;

[0137] 5.5.3) Use the evaluation results to update the Gaussian process surrogate model and update the estimates of the mean and variance;

[0138] 5.5.4) Reselect the next configuration based on the posterior probability estimate.

[0139] 5.6) Convergence and Selection:

[0140] When the predetermined number of iterations is reached or the performance converges, the hyperparameter configuration with the maximum estimated value in the surrogate model is selected as the optimal hyperparameter combination finally found.

[0141] 6) Based on the optimal hyperparameters corresponding to different components, the CNN-BiLSTM combined prediction model described in step 4) is used to perform short-term predictions on different components to obtain predicted values ​​of different components;

[0142] 7) Accumulate the predicted values ​​of different components to obtain the final point prediction result of the daily photovoltaic output (see attached Figure 4 The specific steps of constructing the prediction interval using the Bootstrap method are as follows:

[0143] 7.1) Accumulating the point prediction results obtained for each component in step 6) to obtain the final photovoltaic output point prediction result;

[0144] 7.2) Subtract the actual photovoltaic output data from the photovoltaic output point prediction data to obtain the final photovoltaic output point prediction error data set;

[0145] 7.3) Resample 1000 to 5000 times, that is, randomly extract data points from the prediction error dataset with replacement, repeat 1000 to 5000 times, and create multiple new datasets with the same size as the original data;

[0146] 7.4) Calculate statistics. For each new data set in 7.3), calculate its statistics, such as mean, median, etc.

[0147] 7.5) Constructing a statistical distribution: By repeating steps 7.3) and 7.4), a series of different statistical values ​​are formed, and then a statistical distribution is constructed.

[0148] 7.6) Calculate the confidence interval: Use the statistical distribution calculated in 7.5) and the set confidence level to calculate the final prediction interval.

[0149] 8) Use relevant evaluation indicators to evaluate the point prediction results and interval prediction results respectively.

[0150] 8.1) The evaluation indicators of point prediction results include mean absolute error (MAE), root mean square error (RMSE), goodness of fit (R 2). Among them, MAE reflects the actual error level of the model, RMSE represents the impact of extreme outliers in the prediction results, and R 2 It reflects the overall fit between the model prediction results and the true value. The expressions of each evaluation index are as follows:

[0151]

[0152] in: y i , are the predicted value at the i-th moment, the true value at the i-th moment, and the average of the true data, and n is the number of samples.

[0153] 8.2) The evaluation indicators of interval prediction results include prediction interval coverage (PICP), prediction interval average width (PINAW), and comprehensive index (CWC). The larger the PICP and the smaller the PINAW, the lower the comprehensive index CWC, and the better the interval prediction effect. The calculation formula of each evaluation index is:

[0154]

[0155] CWC=PINAW[1+γ(PICP)e -η(PICP-μ) ] (twenty one)

[0156]

[0157] Where: n is the number of samples; τ i is the Boolean value of the i-th sample. If the predicted value falls within the prediction interval, then τ i Take 1, otherwise take 0. δ is the range of the predicted target, that is, the difference between its maximum and minimum values, used for PINAW normalization. i (x i ) and L i (x i ) are the upper and lower boundaries of the prediction interval respectively. i is the input variable; γ(·) is the penalty coefficient. η is the confidence penalty coefficient of PICP. When PICP does not meet the confidence μ requirement, the comprehensive index CWC will be amplified due to the difference between the PICP value and μ. When CWC meets the confidence μ requirement, the CWC value is determined by PINAW and η. In this paper, η is taken as 50 (see Appendix Figure 5 shown).

[0158] The present invention adopts the Bootstrap method to reconstruct the error sample set of photovoltaic output point prediction, which can realize the interval prediction of photovoltaic output without any assumptions, especially for time series data with large volatility such as photovoltaic output, and has a good prediction effect.

[0159] In order to further illustrate the effectiveness of a short-term photovoltaic output interval prediction method of the present application, the following example is given in combination with the short-term photovoltaic output interval prediction method described in the present application for illustration:

[0160] 1) See Table 1:

[0161] Table 1 Evaluation index values ​​of point prediction results

[0162]

[0163] The effectiveness of the proposed method is verified based on the photovoltaic output data of a photovoltaic power station in Xinjiang Uygur Autonomous Region, China in the spring of 2019, i.e., from March to May. The dataset includes NWP meteorological characteristic data, i.e., temperature, air pressure, humidity, total radiation, direct radiation, and scattered radiation, as well as the component temperature of photovoltaic power generation devices and photovoltaic output history data. The sampling period of the dataset is 15 minutes, 96 sets of data are generated every day, and a total of 8832 samples are included from March to May.

[0164] 2) Linear interpolation was performed on the missing data in the original data, and bad data was automatically screened and removed. The data in the original photovoltaic output data during the period of zero photovoltaic output at night were removed, and only the valid photovoltaic output data of 56 sampling points in the period of 7:00-20:45 were retained. At this point, the total sample size was reduced from the original 8832 to 5152.

[0165] 3) The ICEEMDAN algorithm is used to perform modal decomposition on the photovoltaic output data. The Nstd of ICEEMDAN is set to 0.5, N to 500 times, and M to 2000 times. Finally, 12 modal components including the Res residual component are obtained.

[0166] 4) Use the MIC method to perform feature screening on each modal component, with 0.6 as the threshold, and remove all features with MIC correlation less than 0.6, and only retain features with correlation values ​​higher than 0.6. The features to be screened include historical modal component data, NWP meteorological characteristic data, and temperature characteristics of photovoltaic power generation device components. Among them, the historical modal component data include the value of the modal component 24 hours ago, the value of the modal component 48 hours ago, and the value of the modal component one week ago. The NWP meteorological characteristic data include six features: temperature, air pressure, humidity, total radiation, direct radiation, and scattered radiation.

[0167] 5) For each modal component, the input features obtained by filtering in step 4) are used as the input data set, which is divided into a training set and a test set in a ratio of 8:2. The training set is input into the CNN-BiLSTM model in the form of a sliding window, and the Bayesian optimization algorithm is used to optimize the hyperparameters of the CNN-BiLSTM. The hyperparameters to be optimized for the CNN layer are: the number of channels of the first layer convolution kernel C1, the number of channels of the second layer convolution kernel C2, the size of the first layer pooling layer P1, and the size of the second layer pooling layer P2. The hyperparameters to be optimized for the BiLSTM are: the number of neuron units in the hidden layer of the first layer BiLSTM layer N1, the number of neuron units in the hidden layer of the second layer N2, and the optimization range of the learning rate l of the BiLSTM are all [2,5]. The optimization ranges of C1 and C2 are both [2,5], the optimization ranges of P1 and P2 are both [1,4], the optimization ranges of N1 and N2 are both [16,256], and the optimization range of l is [10 -6 ,0.1]. The number of convolution kernel channels of the first and second layers of CNN are set to 64 and 128 respectively, and MSE Loss is selected as the prediction model loss function of BiLSTM. The optimizer uses Adam optimizer, the number of iterations is set to 300, and the early termination strategy is adopted during the iteration process. If the loss function is not optimized after more than 30 iterations, the hyperparameter optimization is terminated in advance.

[0168] 6) For each modal component, the optimal hyperparameters of CNN-BiLSTM obtained by Bayesian algorithm optimization are used to make a short-term prediction of the modal component value for the next day, and the predicted values ​​of all modal components for the next day are accumulated to obtain the point prediction result of photovoltaic output for the next day. A day in the test set is randomly selected for prediction, and the results are shown in the attached figure. Figure 4 At the same time, the evaluation index of the point prediction result of the day is calculated and shown in Table 1. After the PV output curve with strong volatility is stabilized by the ICEEMDAN algorithm, the change trend of each component has become more obvious, making the CNN-BiLSTM combined prediction model easier to predict. In addition, the Bayesian optimization algorithm finds suitable hyperparameters for the CNN-BiLSTM model of each modal component, making the prediction result very good.

[0169] 7) Based on the point prediction results of photovoltaic output, the prediction error data set was resampled 2000 times using the Bootstrap algorithm, and the confidence intervals at three confidence levels of 95%, 90%, and 85% were calculated to obtain the final prediction interval of photovoltaic output, as shown in the attached figure. Figure 5 At the same time, the evaluation index of the interval prediction result of the day is calculated, as shown in Table 2:

[0170] Table 2 Evaluation index values ​​of interval prediction results

[0171]

[0172] Embodiment 2:

[0173] To achieve the above objectives, the present application also provides a short-term photovoltaic output interval prediction model, which is used to implement the short-term photovoltaic output interval prediction method as described above.

[0174] Embodiment three:

[0175] The present application also provides a computer-readable storage medium, wherein the computer-readable storage medium stores computer-executable instructions, and when the computer-executable instructions are executed by a control processor, the short-term photovoltaic output interval prediction method as described above is implemented.

[0176] The processor may be a central processing unit (CPU), or other general-purpose processors, digital signal processors (DSP), application-specific integrated circuits (ASIC), field-programmable gate arrays (FPGA) or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. A general-purpose processor may be a microprocessor or the processor may be any conventional processor, etc. The memory is used to store an operating system, an application program, a boot loader, data, and other programs, such as the program code of the computer program, etc. The memory may also be used to temporarily store data that has been output or is to be output.

[0177] In several embodiments provided in the present application, it is understood that each box in the flow chart or block diagram can represent a module, a program segment or a part of a code, and the module, a program segment or a part of a code contains one or more executable instructions for realizing the specified logical function. It should also be noted that in some alternative implementations, the functions marked in the box can also occur in a different order from that marked in the accompanying drawings. For example, two consecutive boxes can actually be executed substantially in parallel, and they can sometimes be executed in the opposite order, depending on the functions involved.

[0178] If the functions are implemented in the form of software function modules and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present application can be essentially or partly embodied in the form of a software product that contributes to the prior art. The computer software product is stored in a storage medium and includes several instructions for a computer device to perform all or part of the steps of the methods described in each embodiment of the present application. The aforementioned storage medium includes: various media that can store program codes, such as a USB flash drive, a mobile hard disk, a read-only memory (ROM), a random access memory (RAM), a magnetic disk or an optical disk.

[0179] The specific embodiments described above further describe the purpose, technical solutions and beneficial effects of the present application in detail. It should be understood that the above description is only a specific embodiment of the present application and is not intended to limit the scope of protection of the present application. It is particularly pointed out that for those skilled in the art, any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present application should be included in the scope of protection of the present application.

Claims

1. A short-term photovoltaic output interval prediction method, characterized in that: The main steps include: S1: Obtain the photovoltaic output curve data of the photovoltaic power station within time T, pre-process the original data, and obtain a real data set; S2: performing modal decomposition on the real data set based on an improved adaptive noise complete set empirical mode decomposition algorithm to obtain multiple modal components; S3: using the maximum mutual information coefficient method to perform feature screening on the multiple modal components respectively, and screen out optimal input feature data for different modal components; S4: constructing different CNN-BiLSTM combination prediction models based on the multiple modal components and their optimal input feature combinations; S5: Using the Bayesian optimization algorithm to optimize the hyperparameters of the CNN-BiLSTM combined prediction model to obtain the optimal hyperparameters; S6: Based on the optimal hyperparameters corresponding to different components, the CNN-BiLSTM combined prediction model is used to perform short-term photovoltaic output prediction for different components to obtain photovoltaic output prediction values ​​for different components; S7: Accumulating the photovoltaic output prediction values ​​of the different components to obtain a point prediction result of the final daily photovoltaic output, and reconstructing its prediction error based on the Bootstrap algorithm to obtain a final photovoltaic output prediction interval.

2. A short-term photovoltaic output interval prediction method according to claim 1, characterized in that: The step S2 specifically includes: S21: Add a set of white noise ω to the original signal x(t) i (t), construct the sequence x i (t), and obtain the first group of residual components δ1(t): x i (t)=x(t)+γ0E1(ω i (t)); δ1(t)=<M(x1(t))> ; Where x(t) represents the original signal to be decomposed; ω i (t) indicates the i-th time Gaussian white noise is added; x i (t) is the new signal constructed after adding white noise for the ith time; γ k-1 is the adaptive coefficient for solving the k-th order modal component; E k (·) represents the kth-order modal component generated by EMD decomposition; δ k represents the k-th order residual component obtained by ICEEMDAN; M(·) represents the local mean of the signal; <·> represents the overall mean; S22: Calculate the first-order modal component f1: f1=x(t)-δ1; S23: Continue to add white noise and express the second set of residual components δ2(t) as δ1+γ1E2ω i (t) mean, and obtain the second-order modal component f2: f2=δ1-δ2=δ1- <M(δ1+γ1E2ω i (t))>; S24: Repeat step S23 to obtain the k-th order residual component δ k With the modal component f k : d k = <M(δ k-1 +g k-1 E k oh i (t))>; f k =d k-1 -d k =d k-1 - <M(δ k-1 +g k-1 E k oh i (t))>; S25: until all modal components and residual components are obtained, the modal decomposition is completed.

3. A short-term photovoltaic output interval prediction method according to claim 1, characterized in that: The step S3 specifically includes: S31: For a binary data set C(x,y), divide it into a and b intervals in the X direction and Y direction respectively, then a×b grids G are formed in the two-dimensional plane. Let the set of grids G formed by different division methods be Ω, C|G represents the distribution of the data set C on the network G, then: In the formula, I * (C,a,b) is the maximum mutual information of X and Y under data set C; S32: The feature matrix composed of the maximum normalized mutual information of the data set C is represented as M, then the element m ab Defined as: S33: The final MIC value is: Where N is the total number of samples; C(N) is the sample number function, generally C(N) = N 0.6 ; S34: MIC correlation values ​​between different features and different modal components are calculated according to S31-S33, a threshold is set, features with MIC correlation values ​​less than the threshold are eliminated, and the remaining features are used as the optimal input feature data corresponding to different modal components.

4. A short-term photovoltaic output interval prediction method according to claim 1, characterized in that: The step S4 specifically includes: S41: Dividing the optimal input feature data of the different modal components obtained in step S2 and the different components obtained in step S3 into a training set and a test set according to a certain ratio; S42: Constructing convolutional neural network layers in sequence: convolutional layer, pooling layer, convolutional layer, and pooling layer. The training set is constructed as a feature map in the form of a sliding window and input into the convolutional neural network layer. The hidden layer is used to extract feature vectors that can characterize photovoltaic output information. S43: construct a BiLSTM network layer, which is composed of forward and reverse LSTM neural networks; the feature vector extracted by the convolutional neural network layer will be input into the BiLSTM neural network after being flattened by the Flatten layer, and the BiLSTM captures the bidirectional time series correlation of the photovoltaic output data; wherein, the specific expression of the BiLSTM neural network is as follows: Where: x t , A t , B t and t represents the input data at time t, the hidden layer of the forward iteration, the hidden layer of the backward iteration, and the output data. w1, w2, …, w6 are the weights of the corresponding layers, and η1, η2, η3 are the activation functions of the corresponding layers.

5. A short-term photovoltaic output interval prediction method according to claim 1, characterized in that: The step S5 specifically includes: S51: Determine hyperparameter space: Define the hyperparameter space of the CNN-BiLSTM model; S52: Select Gaussian process as the proxy model: initialize the Gaussian process proxy model as the probability model of the objective function; S53: Adopting Gaussian process posterior maximization strategy as the selection sampling strategy to select the next sample point in the hyperparameter space to evaluate the objective function; S54: Randomly initialize the proxy model and the sampling strategy; S55: Iterate and optimize to obtain the optimal hyperparameter combination.

6. A short-term photovoltaic output interval prediction method according to claim 5, characterized in that: The step S55 specifically includes: S551: Based on the posterior probability maximization strategy of the Gaussian process surrogate model, select the next hyperparameter configuration to maximize the potential performance improvement; S552: Evaluate the configuration on the real model to obtain a performance indicator, i.e., an error in photovoltaic output prediction; S553: ​​Use the evaluation results to update the Gaussian process surrogate model and update the estimates of the mean and variance; S554: reselect the next configuration based on the posterior probability estimate; S555: When a predetermined number of iterations is reached or the performance converges, the hyperparameter configuration with the maximum estimated value in the proxy model is selected as the optimal hyperparameter combination finally found.

7. A short-term photovoltaic output interval prediction method according to claim 1, characterized in that: The step S7 specifically includes: S71: Accumulating the photovoltaic output prediction values ​​of different components in step S6 to obtain the final photovoltaic output point prediction result; S72: subtracting the actual photovoltaic output data from the photovoltaic output point prediction result to obtain a final photovoltaic output point prediction error data set; S73: resampling to create multiple new data sets; wherein the size of the data sets is the same as the original data; S74: Calculate statistics, for each new data set in S73, calculate its statistics; S75: repeating steps S73 and S74 to form a series of different statistical values ​​and construct a distribution of the statistics; S76: Using the distribution of the statistic calculated in S75 and according to the set confidence level, a final photovoltaic output prediction interval is calculated.

8. A short-term photovoltaic output interval prediction model, characterized in that: The model is used to implement the short-term photovoltaic output interval prediction method as described in any one of claims 1-7.

9. A computer-readable storage medium storing computer-executable instructions, characterized in that: When the computer executable instructions are executed by the control processor, the short-term photovoltaic output interval prediction method as described in any one of claims 1-7 is implemented.

Images