Ultra-short-term photovoltaic power generation power prediction method based on double-layer decomposition and deep learning nonlinear integrated model

Abstract

The invention provides an ultra-short-term photovoltaic power generation power prediction method based on a double-layer decomposition and deep learning nonlinear integrated model, and the method comprises the steps: carrying out the data cleaning of historical photovoltaic power data and meteorological data, and screening the meteorological feature data with strong correlation after the data cleaning; decomposing the historical photovoltaic power data after data cleaning into a plurality of sub-components by using an ICEEMDAN algorithm, calculating an SE value of each sub-component, performing modal reconstruction on the sub-components with similar complexity according to the SE to obtain high-frequency components, periodic components and low-frequency components with different scales, and performing secondary decomposition on the high-frequency components by using a VMD algorithm; respectively constructing a prediction model based on BiLSTM-CNN-Attention for the sub-components, training the prediction model by using the sample data set, and outputting a prediction result of each prediction model; and performing nonlinear integration on prediction results of the sub-components by using a random forest to obtain a final power prediction value. According to the method, a power prediction value with relatively high progress can be obtained.

Description

Technical Field

[0001] This application belongs to the field of new energy power generation technology, specifically, it relates to an ultra-short-term photovoltaic power generation prediction method based on a two-level decomposition and a deep learning nonlinear ensemble model. Background Technology

[0002] In recent years, with the depletion of fossil fuels and changes in the international situation, the transformation of the global energy structure has accelerated, and the development and utilization of renewable energy has become a primary task for countries to achieve green recovery. Solar energy, as an important form of sustainable energy, has gained widespread global recognition for its substitutive benefits. However, with the emergence of numerous grid-connected photovoltaic (PV) systems, the intermittency and randomness of PV power generation have become increasingly prominent, posing a significant challenge to the stable operation of the power grid and economic planning. Accurate PV power generation forecasting helps mitigate the losses to the power grid caused by these problems, thereby improving the operational efficiency of PV power plants and contributing to a high-quality and stable power supply to the grid. Therefore, accurate ultra-short-term PV power generation forecasting is crucial for stable grid operation and optimal energy allocation.

[0003] Since the timescale of ultra-short-term forecasts is generally between 0 and 4 hours, it is easier to achieve accurate photovoltaic power forecasts, thereby determining the grid market's dispatch situation and planning schemes.

[0004] Current prediction methods are mainly divided into physical methods, statistical methods, and metaheuristic methods. With the rise of artificial intelligence, metaheuristic algorithms based on machine learning models provide more reliable solutions and cutting-edge technologies for accurate photovoltaic power prediction. Shallow networks based on machine learning, using a single hidden layer, may fail to extract deep nonlinear features of the data. To address this issue, hidden layers are added to enhance the nonlinear mapping capability. However, too many layers increase the complexity of the network structure, potentially leading to serious problems such as vanishing gradients, local minima, and overfitting. Deep learning, with its strong feature transformation capabilities, is an effective method to overcome the shortcomings of shallow models, and is particularly suitable for ultra-short-term photovoltaic power prediction.

[0005] The output power of photovoltaic power generation systems exhibits time-varying and non-stationary characteristics. Effective data preprocessing can reduce the difficulty of prediction. Therefore, combining signal decomposition strategies with deep learning further improves the accuracy of prediction results. However, decomposition still has two problems: First, when inputting the decomposed subsequences into the model for prediction, the complexity and correlation between the subsequences are not considered, thus reducing prediction efficiency and accuracy. Second, after obtaining the prediction results of the subsequences, the final reconstruction is limited to a linear summation. Linear reconstruction is not suitable for all situations and may affect the accuracy of prediction. Summary of the Invention

[0006] The technical problem addressed in this application is: how to provide an ultra-short-term photovoltaic power generation prediction method based on a two-layer decomposition and deep learning nonlinear ensemble model that can improve the prediction accuracy of photovoltaic power generation and reduce the negative impact of the intermittency and randomness of photovoltaic power generation on the power system.

[0007] This application provides a method for predicting ultra-short-term photovoltaic power generation based on a two-level decomposition and a deep learning nonlinear ensemble model. The prediction method includes:

[0008] Data preprocessing stage: Historical photovoltaic power data and meteorological data are cleaned, and highly correlated meteorological feature data are selected after data cleaning;

[0009] Two-level decomposition stage: The ICEEMDAN algorithm is used to decompose the cleaned historical photovoltaic power data into several sub-components, calculate the SE value of each sub-component, and perform mode reconstruction on sub-components with similar complexity based on the SE to obtain high-frequency components, periodic components and low-frequency components of different scales. The VMD algorithm is used to perform secondary decomposition on the high-frequency components.

[0010] Model building phase: Construct prediction models based on BiLSTM-CNN-Attention for each sub-component, and train the prediction models using the sample dataset (historical photovoltaic power data and meteorological data after data preprocessing), and output the prediction results of each prediction model;

[0011] Nonlinear integration stage: The prediction results of the sub-components are nonlinearly integrated using random forest to obtain the final power prediction value.

[0012] This application provides an ultra-short-term photovoltaic power generation prediction method based on bi-level decomposition and a deep learning nonlinear ensemble model, which has the following technical advantages:

[0013] (1) Important meteorological features are selected from both linear and nonlinear perspectives by using Pearson correlation coefficient, Spearman correlation coefficient and MIC coefficient to avoid redundant features from interfering with the prediction model.

[0014] (2) The two-level decomposition method of ICEEMDAN-SE-VMD effectively improves the quality of sample data and reduces the complexity of non-stationary power sequences. Among them, the mode reconstruction method based on SE value not only ensures the similarity complexity between sub-components but also improves the prediction efficiency. In addition, it avoids the superposition error caused by too many sub-components.

[0015] (3) The BiLSTM-CNN-Attention model fully explores deep-level important features based on spatiotemporal correlation analysis, ensuring the model's feature expression ability.

[0016] (4) Using random forest to perform nonlinear ensemble of the predicted values ​​of subsequences effectively expresses the nonlinear relationship between subcomponents and improves the prediction accuracy. Attached Figure Description

[0017] Figure 1 The flowchart shows the main steps of an ultra-short-term photovoltaic power generation prediction method based on a bi-layer decomposition and deep learning nonlinear ensemble model according to one or more embodiments.

[0018] Figure 2 This is a detailed flowchart illustrating an ultra-short-term photovoltaic power generation prediction method based on a two-layer decomposition and deep learning nonlinear ensemble model, according to one or more embodiments.

[0019] Figure 3 This is a two-layer exploded view according to one or more embodiments;

[0020] Figure 4 A network structure diagram of BiLSTM according to one or more embodiments;

[0021] Figure 5 A network structure diagram of CNN-Attention according to one or more embodiments;

[0022] Figure 6 Detailed structure and parameter diagrams of the BiLSTM-CNN-Attention model according to one or more embodiments;

[0023] Figure 7 This is an architecture diagram of a random forest according to one or more embodiments. Detailed Implementation

[0024] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0025] Before describing the various embodiments of this application in detail, the technical concept of this application is first briefly described: Currently, when performing ultra-short-term photovoltaic power prediction, the intermittency and randomness of photovoltaic power generation systems can easily adversely affect the prediction accuracy of power generation. Therefore, the ultra-short-term photovoltaic power prediction method based on bi-level decomposition and a deep learning nonlinear ensemble model provided in this application first performs data cleaning and screening of relevant meteorological data in the data preprocessing stage, followed by bi-level decomposition to reduce the noise level and prediction difficulty of the original power sequence; in the model construction stage, BiLSTM-CNN-Attention networks are constructed for each sub-component, and the model is trained using a sample dataset to output its respective prediction results; in the nonlinear ensemble stage, random forest is used to nonlinearly integrate the prediction results of the sub-components to obtain the final power prediction value. This method utilizes the BiLSTM-CNN-Attention combined model to extract key information of spatiotemporal complementary features and uses random forest to effectively express the nonlinear relationship between sub-components, thereby improving the final prediction accuracy. The specific principles of the ultra-short-term photovoltaic power prediction method based on bi-level decomposition and a deep learning nonlinear ensemble model of this application are described below with more embodiments.

[0026] Specifically, such as Figure 1 and Figure 2 As shown, the ultra-short-term photovoltaic power generation prediction method based on a two-layer decomposition and deep learning nonlinear ensemble model in this embodiment includes:

[0027] Step S10, Data Preprocessing Stage: Clean the historical photovoltaic power data and meteorological data, and select the meteorological feature data with strong correlation after data cleaning.

[0028] Step S20, Two-layer decomposition stage: The ICEEMDAN algorithm is used to decompose the cleaned historical photovoltaic power data into several sub-components, calculate the SE value of each sub-component, and perform mode reconstruction on sub-components with similar complexity based on the SE to obtain high-frequency components, periodic components and low-frequency components of different scales. The VMD algorithm is used to perform secondary decomposition on the high-frequency components.

[0029] Step S30, Model Building Stage: Construct prediction models based on BiLSTM-CNN-Attention for each sub-component, train the prediction models using the sample dataset, and output the prediction results of each prediction model.

[0030] Step S40, Nonlinear Integration Stage: The prediction results of the sub-components are nonlinearly integrated using random forest to obtain the final power prediction value.

[0031] In one or more embodiments, the method for cleaning historical photovoltaic power data and meteorological data includes: filling in missing values ​​in the historical photovoltaic power data and meteorological data; detecting outliers in the historical photovoltaic power data and meteorological data after filling in missing values, and removing outliers.

[0032] For example, firstly, missing value detection is performed on the original dataset (historical photovoltaic power data and meteorological data), and then missing values ​​are imputed by averaging the data before and after the missing values. After imputation, outlier detection is performed on the dataset. Box plots are used to detect outliers in the original dataset, and then the outliers are corrected using the mean of the dataset. The criteria for judging outliers are: the value exceeds the upper quartile plus 1.5 times the interquartile range or is less than the lower quartile minus 1.5 times the interquartile range.

[0033] In one or more embodiments, the method for selecting strongly correlated meteorological feature data after data cleaning includes: simultaneously analyzing the correlation between temperature, relative humidity, wind speed, wind direction, daily rainfall, total horizontal radiation, horizontal scattering, total tilted radiation, and tilted scattering with power generation using Pearson correlation coefficient, Spearman correlation coefficient, and MIC coefficient. This allows for a comprehensive consideration of both linear and nonlinear relationships. Finally, total horizontal radiation, horizontal scattering, total tilted radiation, tilted scattering, temperature, and wind speed are selected as the meteorological input parameters for the model.

[0034] In one or more embodiments, such as Figure 3 As shown, the method for decomposing historical photovoltaic power data after data cleaning into several sub-components using the ICEEMDAN algorithm includes:

[0035] Add i sets of white noise w to the original signal (i) The sequence x is constructed after EMD decomposition. (i) :

[0036] x (i) =x+β0E(w) (i) ), R1=(N(x (i) ))

[0037] In the formula: x is the signal to be decomposed, β represents the white noise energy parameter, and w (i) R1 represents Gaussian white noise, and R1 represents the first set of residuals.

[0038] The first mode component is obtained by subtracting the residual from the original signal:

[0039] d1 = x - R1

[0040] Continue adding white noise, and use local mean decomposition to calculate the second set of residuals R1+β1E(w) (i) ), thus obtaining the second modal component:

[0041] d2 = R1 - R2 = R1 - (N(R1 + β1E(w (i) )))

[0042] Calculate the residual sum and modal component of the k-th group in sequence:

[0043] R k = (N(R k-1 + β k-1 E(w (i) )))

[0044] d k = R k-1 - R k

[0045] Repeat this step until all the modes and residuals are obtained, and the decomposition ends.

[0046] Exemplarily, the original power adaptivity is decomposed into 9 eigenmode components and one residual component. IMFs of different scales contain different characteristic information. The high-frequency components reflect short-term random information; the low-frequency components reflect the long-term trend of the power. From IMF1 to Res, the modal oscillation frequency gradually decreases and the oscillation period gradually increases. Compared with the original sequence, the fluctuation law of the sub-components is more stable and regular.

[0047] Exemplarily, the method for calculating the SE value of each sub-component includes:

[0048] Construct a time series {x i} = {x1, x2, …, x N} of length N into an m-dimensional vector:

[0049] X i = {x i , x i+1 , …, x i+m-1}, (i = 1, 2, …, N - m + 1)

[0050] Define the distance d m [X(i), X(j)] (i ≠ j) between vector X(i) and X(j) as the absolute value of the maximum difference between their corresponding elements:

[0051]

[0052] Given a threshold r (r > 0), count the number of d m [X(i), X(j)] < r for each i value, denoted as N m (i); and calculate the ratio to the total number of vectors N - m, denoted as

[0053]

[0054] All obtained by solving Calculate the average:

[0055]

[0056] Update m to m+1, repeat the above steps, and obtain B. m+1 (r):

[0057]

[0058] The calculated estimated value of SE is:

[0059]

[0060] Based on the entropy calculation results of the sub-components, sub-components with similar SE values ​​are reconstructed to obtain high-frequency components, periodic components, and low-frequency components.

[0061] For example, a method for performing secondary decomposition of high-frequency components using the VMD algorithm includes:

[0062] Performing a Hilbert transform on the initial signal yields an analytic signal with k modal components, along with its one-sided spectrum.

[0063]

[0064] In the formula: δ(t) is the impulse function;

[0065] Each mode u k The spectrum of (t) is modulated onto the fundamental frequency band to obtain:

[0066]

[0067] By estimating the bandwidth of each mode signal through the Gaussian smoothness of the demodulated signal, the variational problem is constructed as follows:

[0068]

[0069] In the formula: {u k} = {u1, ..., u k} represents the k-th mode; {ω k}={ω1,…,ω k} represents u k The kth center frequency;

[0070] Transforming the aforementioned constraint into an unconstrained variational problem, and introducing a quadratic penalty factor α and a Lagrange multiplication operator λ, the extended Lagrange expression is:

[0071]

[0072] The solution is obtained using the alternating direction method of multiplication operators, and u is updated. k and ω k :

[0073]

[0074] In the formula: and They are respectively f(ω) and λ n Fourier transform of (ω), where n is the number of iterations.

[0075] For example, VMD is used to decompose the high-frequency components into 6 subsequences. Clearly, the fluctuation trend of the IMFs is more regular compared to the high-frequency components, clearly expressing the contained feature information, which is beneficial for further exploring the potential fluctuation characteristics of the high-frequency components.

[0076] In one or more embodiments, such as Figure 4 As shown, the BiLSTM network of the prediction model includes an input layer, an LSTM forward layer, an LSTM backward layer, and an output layer. The input layer is used to input the sequence, the LSTM forward layer is used to process the input sequence in the forward direction, the LSTM backward layer is used to process the input sequence in the reverse direction, and the output layer is used to output the prediction result. Specifically, the BiLSTM network is composed of two LSTMs, one processing the input sequence in the forward direction and the other processing the sequence in the backward direction, thereby obtaining the information flow of the entire time domain. The bidirectional network enables recursive feedback to the hidden layers of the past and future, further improving the utilization of feature data. The specific steps are as follows:

[0077] f t =σ(W f ·[h t-1 ,X t ]+b f )

[0078] i t =σ(W i ·[h t-1 ,X t ]+b i )

[0079]

[0080] O t =σ(W O ·[h t-1 ,X t ]+b O )

[0081] h t =O t⊙tanh(C t )

[0082]

[0083] In the formula: f t For the Gate of Oblivion, i t For input gate, O t For output gate, W f W i W C and W O Let b be the weight matrix. f b i b C and b O Let σ be the bias vector, σ be the sigmoid activation function, and tanh be the hyperbolic tangent activation function. For candidate cell states, C t For the new memory cell state, h t This represents the output of the hidden layer, ⊙ represents the product of corresponding points, and LSTM represents the LSTM network function. Indicates the state of the forward hidden layer. U represents the reverse hidden layer state. t and V t These are the output weights of the forward LSTM propagation unit and the backward LSTM propagation unit, respectively, b y This is the bias vector of the output layer at the current time.

[0084] In one or more embodiments, the CNN-Attention network is a feedforward neural network characterized by parameter sharing and sparse connections. A typical CNN mainly consists of convolutional layers, pooling layers, and fully connected layers. The convolution operation is as follows:

[0085]

[0086] In the formula: k represents the convolution kernel of the convolutional layer, x is the input signal, b is the bias term of the corresponding convolutional layer, and f is the nonlinear activation function ReLU.

[0087] Pooling layers include max pooling and average pooling. In order to improve the CNN's attention to important information and suppress irrelevant information from interfering with the model, this embodiment adds an Attention layer to the CNN. Its core idea is to quickly filter key features by assigning different weights to feature vectors.

[0088] For example, such as Figure 5 and Figure 6As shown, the CNN-Attention network of the prediction model consists of an input layer, a convolutional layer, a pooling layer, another convolutional layer, another pooling layer, an attention mechanism layer, a fully connected layer, and an output layer, connected sequentially. The prediction model based on BiLSTM-CNN-Attention is an end-to-end stacked network. The front-end BiLSTM extracts long-term temporal features, while the back-end CNN-Attention captures key information from these long-distance temporal features, fully exploiting time-invariant spatial local features. Based on the spatiotemporal correlation analysis of these features, it completes the ultra-short-term prediction task of photovoltaic power. The BiLSTM-CNN-Attention network consists of 13 layers. First, multivariate features are passed to the input layer of the BiLSTM in matrix form, with tanh as the activation function. To prevent overfitting, a Dropout layer is added to the BiLSTM. The features extracted by the BiLSTM are concatenated to Conv1D for convolutional processing. Then, the feature map formed by the convolution operation is fed to a max-pooling layer for downsampling, reducing the dimensionality of the data. Adding a Batch Normalization (BN) layer after each convolutional layer helps reduce noise levels during training. The features after convolution and pooling are input into the Attention layer, where the softmax function assigns attention weights to the feature vectors. The important features extracted by CNN-Attention are then fused using the Flatten layer. Finally, the prediction result is obtained through a fully connected layer and an output layer. In the experiments, the optimizer was Adam, the learning rate was set to 0.0001, the number of iterations was set to 256, and the batch size was set to 512.

[0089] In one or more embodiments, such as Figure 7 As shown, the method of nonlinearly integrating the prediction results of sub-components using random forest to obtain the final power prediction value includes:

[0090] k sample sets are randomly selected from the original sample set (the prediction values ​​of each model obtained during the model building phase), and each sample set contains M training samples.

[0091] Training is performed on k sample sets respectively to form k corresponding weak learners;

[0092] When training nodes in a decision tree, a subset of samples is selected from all samples at each node, and an optimal feature is chosen from these subset features based on the Gini coefficient to split the left and right subtrees. The standard definition of the Gini coefficient is:

[0093] Gini(e) = 2e - 2e 2

[0094] The final model output is obtained by arithmetically averaging the predicted regression results obtained from k weak learners.

[0095] For example, the ultra-short-term prediction duration of the prediction method in this embodiment is 1 hour.

[0096] The specific embodiments of this application have been described in detail above. Although some embodiments have been shown and described, those skilled in the art should understand that modifications and improvements can be made to these embodiments without departing from the principles and spirit of this application as defined by the claims and their equivalents, and such modifications and improvements should also be within the protection scope of this application.

Claims

1. A method for ultra-short-term photovoltaic power prediction based on a two-layer decomposition and deep learning nonlinear ensemble model, characterized in that, The prediction method comprises: a data preprocessing stage: data cleaning is performed on historical photovoltaic power data and meteorological data, and meteorological feature data with strong correlation is screened out after data cleaning; a double-layer decomposition stage: the historical photovoltaic power data after data cleaning is decomposed into a plurality of sub-components by using an ICEEMDAN algorithm, SE values of the sub-components are calculated, sub-components with similar complexity are reconstructed according to SE, high-frequency components, periodic components and low-frequency components of different scales are obtained, and the high-frequency components are decomposed again by using a VMD algorithm; a model construction stage: a prediction model based on BiLSTM-CNN-Attention is constructed for the sub-components, the prediction model is trained by using a sample data set, and a prediction result of each prediction model is output; a nonlinear integration stage: the prediction results of the sub-components are nonlinearly integrated by using a random forest, and a final power prediction value is obtained.

2. The method of Claim 1, wherein the method is characterized by, The method for performing data cleaning on historical photovoltaic power data and meteorological data comprises: missing value filling is performed on the historical photovoltaic power data and the meteorological data; abnormal value detection is performed on the historical photovoltaic power data and the meteorological data after the missing value filling, and the abnormal values are removed.

3. The method of Claim 2, wherein the method is characterized by, The method for screening out meteorological feature data with strong correlation after data cleaning comprises: Pearson correlation coefficient, Spearman correlation coefficient and MIC coefficient are used to analyze the correlation between temperature, relative humidity, wind speed, wind direction, daily rainfall, total horizontal radiation, horizontal scattering, total inclined radiation, inclined scattering and photovoltaic power; total horizontal radiation, horizontal scattering, total inclined radiation, inclined scattering, temperature and wind speed are selected as the meteorological feature data.

4. The method of Claim 3, wherein the method is characterized by, The method for decomposing the historical photovoltaic power data after data cleaning into a plurality of sub-components by using the ICEEMDAN algorithm comprises: adding i sets of white noise w to the original signal (i) , constructing a sequence x after EMD decomposition (i) : x (i) = x + β0E(w (i) ), R1= (N(x (i) )) In the formula: x is the signal to be decomposed, β represents a white noise energy parameter, w (i) represents a Gaussian white noise, and R1 represents the first set of residuals. a first modal component is obtained by subtracting a residual from an original signal: d1=x-R1 Continuing to add white noise, the second set of residuals R1+β1E(w (i) ) is calculated using local mean decomposition, and the second modal component is obtained: d2 = R1 - R2 = R1 - (N(R1 + β1E(w (i) ))) a kth residual and a modal component are calculated in sequence: R k = (N(R k-1 + β k-1 E(w (i) ))) d k = R k-1 - R k the step is repeated until all modalities and residual numbers are obtained, and the decomposition is completed.

5. The method of claim 3, wherein the method is characterized by, The method for decomposing the high-frequency components again by using the VMD algorithm comprises: Hilbert transformation is performed on an initial signal to obtain an analytic signal of k modal components, and a one-sided frequency spectrum is obtained simultaneously: wherein, δ(t) is an impulse function; The modals u k (t) are spectrally modulated onto the base band, resulting in: a bandwidth of each modal signal is estimated by Gaussian smoothing of the demodulated signal, and a variational problem is constructed as follows: In the formula: {u k } = {u1, ..., u k } represents the k-th mode; {ω k }={ω1,…,ω k } represents u k The kth center frequency; the above constraint is converted into a non-constraint variational problem, a quadratic penalty factor α and a Lagrange multiplier λ are introduced, and an extended Lagrange expression is: Solve by using the multiplicative alternating direction method, and update u k and ω k : wherein: and are respectively f the Fourier transform of (ω) and λ n (ω), n is the number of iterations.

6. The method of claim 1, wherein the method is characterized by, The method for calculating SE values of the sub-components comprises: A time series {x i} of length N is constructed into an m-dimensional vector: N} = {x1,x2,…,x X i = {x i ,x i+1 ,…,x i+m-1}, (i = 1, 2, …, N-m+1) The distance d between vectors X(i) and X(j) is defined as m [X(i), X(j)] (i≠j) is the absolute value of the largest difference between the corresponding elements of the two vectors. For a given threshold r (r > 0), count d based on each i value m The number of [X(i), X(j)] < r, denoted as N m (i); and calculate the ratio to the total number of vectors N - m, denoted as All of the solutions obtained are Average: updating m as m+1, repeating the above steps to obtain B m+1 (r): an estimated value of SE is calculated as:

7. The method of claim 1, wherein the method is characterized by, The BiLSTM network of the prediction model comprises: an input layer for inputting a sequence; an LSTM forward layer for forward processing of the input sequence; an LSTM reverse layer for reverse processing of the input sequence; an output layer for outputting a prediction result.

8. The method of claim 7, wherein the method is characterized by, The CNN-Attention network of the prediction model comprises an input layer, a convolution layer, a pooling layer, a convolution layer, a pooling layer, an attention mechanism layer, a full connection layer and an output layer connected in sequence. 9.The ultra-short-term photovoltaic power forecasting method based on the two-layer decomposition and deep learning nonlinear ensemble model according to claim 1, wherein, The method for nonlinearly integrating the prediction results of the sub-components by using the random forest to obtain the final power prediction value comprises the following steps: randomly sampling k sampling sets from the original sample set, each sampling set containing M training samples; training the k sampling sets respectively to form corresponding k weak learners; when training the decision tree nodes, selecting part of the samples from all the samples of the node, and selecting an optimal feature in the part of the samples according to the Gini coefficient to split the left and right sub-trees; performing arithmetic averaging on the prediction regression results obtained by the k weak learners to obtain the final model output.

10. The method of claim 1 to 9, wherein the method is characterized in that, The prediction method has a super-short-term prediction duration of 1 hour.

Images