A photovoltaic power generation prediction method based on NGBoost algorithm and energy storage control

Abstract

The present invention discloses a photovoltaic power generation prediction method based on the NGBoost algorithm and energy storage control, which comprises the following steps: dividing the collected original data set into multiple groups of feature columns, performing outlier detection on each feature column; replacing the outliers to obtain a preprocessed data set; decomposing the preprocessed light intensity into multiple modules, and then dividing the data of each module into a point training set and a point test set, training and evaluating the A-BiGRU point prediction model to obtain point predicted light intensity; performing interval prediction on the predicted light intensity using the NGBoost interval prediction model to obtain interval predicted light intensity; converting the light intensity into photovoltaic power generation to obtain predicted photovoltaic power generation and photovoltaic power generation corresponding to the upper and lower limits of the interval; selecting an appropriate energy storage capacity based on the power generation corresponding to the upper and lower limits of the interval, and constructing an energy storage control prediction system. The present invention improves the prediction accuracy and meets the requirements for grid connection.

Description

Technical Field

[0001] The present invention relates to the field of power electronics technology, and in particular to a photovoltaic power generation prediction method based on an NGBoost algorithm and energy storage control. Background Art

[0002] As a new energy source, solar energy has attracted more and more attention due to its huge storage capacity and cleanliness. Photovoltaic power generation has become the main trend of current new energy power generation and has attracted attention from countries around the world. Large-scale use of photovoltaic power generation has also become the preferred approach.

[0003] However, photovoltaic power generation is affected by many factors, and its output is indirect, random, and volatile. This can impact the security and stability of the power grid during the grid connection process. Therefore, reasonable and high-precision photovoltaic power generation forecasts are necessary. Currently, photovoltaic power generation forecasts are mainly calculated using methods such as physics, statistics, and machine learning.

[0004] Physical models have a narrow scope of application and are overly limited in their use conditions. They are not effective in predicting solar irradiance under drastic weather conditions or over long time scales. Statistical methods have high requirements for data volume and quality, requiring large amounts of high-quality data for training to achieve ideal prediction accuracy, which is greatly limited. Single machine learning methods are often affected by the model's own strengths and weaknesses, and their prediction accuracy is often low. Solar intensity is highly random and its value is affected by many factors, resulting in strong uncertainty, making the fitting effect of a single model unsatisfactory. Therefore, the current main research direction is to use artificial intelligence algorithms based on hybrid models for prediction. However, even with the use of hybrid models, the prediction accuracy of these methods is currently difficult to further improve significantly and still falls short of the actual accuracy requirements. To further improve the model's accuracy, more effective machine learning models are needed, or the model architecture is adjusted according to local conditions. The former is more difficult and difficult to achieve in a short period of time. Although the latter may perform well under specific conditions, it is difficult to maintain high accuracy when the application scenario changes. The model architecture needs to be adjusted again, and it can only be put into use after multiple attempts, which undoubtedly reduces the versatility of the method and increases unnecessary time costs. Summary of the Invention

[0005] Purpose of the invention: The purpose of the present invention is to provide a photovoltaic power generation prediction method based on the NGBoost algorithm and energy storage control, so as to improve the accuracy of photovoltaic power generation prediction and meet the photovoltaic grid-connected accuracy requirements.

[0006] Technical solution: A photovoltaic power generation prediction method based on the NGBoost algorithm and energy storage control, including the following steps:

[0007] S1, the collected original data set is divided into multiple groups of feature columns according to time and features, and the isolation forest method is used to detect outliers for each feature column;

[0008] S2, uses the cubic spline interpolation method to replace the outliers in each feature column, and then merges all feature columns to obtain the preprocessed data set;

[0009] S3, using the CEEMDAN frequency division algorithm, decomposes the preprocessed light intensity into multiple modules; then divides the data of each module into a point training set and a point test set, trains and evaluates the A-BiGRU point prediction model respectively, and saves the best A-BiGRU point prediction model; uses meteorological data X as the input of the A-BiGRU point prediction model, linearly adds the point prediction results of each module, and obtains the point prediction light intensity result Y i ;

[0010] S4, using the preprocessed data set and the point predicted light intensity result, performs interval prediction on the predicted light intensity through the NGBoost interval prediction model to obtain the interval predicted light intensity result;

[0011] S5, using the MPPT algorithm to convert the light intensity into photovoltaic power generation, and obtain the predicted photovoltaic power generation and the photovoltaic power generation corresponding to the upper and lower limits of the light intensity range;

[0012] S6, select the appropriate energy storage capacity according to the upper and lower limits of power generation, and build an energy storage control prediction system; set the energy storage control logic as follows: when the actual power generation is greater than the predicted power generation, the energy storage is charged; when the actual power generation is less than or equal to the predicted power generation, the energy storage is discharged.

[0013] Furthermore, the original data set includes historical meteorological data and light intensity data, and the features include humidity, temperature, wind direction, and air pressure. The steps for implementing outlier detection for each feature column using the isolation forest method are as follows:

[0014] S11, put n samples of a single feature column into the root node of the tree;

[0015] S12, the isolation forest algorithm randomly specifies a dimension and randomly generates a cut point in the current node data. The cut point is generated between the maximum and minimum values ​​of the specified dimension in the current node data;

[0016] S13: Generate a hyperplane based on the cutting point and divide the data space of the current node into two subspaces: put the data less than p in the specified dimension into the left leaf of the current node, and put the data greater than or equal to p into the right leaf of the current node;

[0017] S14, looping steps S12 and S13 in the leaf nodes, continuously constructing new leaf nodes until each leaf node contains only one data, and obtaining n isolation forests;

[0018] S15, repeat steps S11-S14, and filter out the outliers of each feature column using the isolation forest method.

[0019] Furthermore, in step S2, the implementation steps of replacing the outliers of each feature column using the cubic spline interpolation method are as follows:

[0020] S21, for a single feature column, record the sequence label corresponding to the outlier and then remove the outlier;

[0021] S22, constructing a cubic spline interpolation function for the feature column after removing outliers;

[0022] S23, using the sequence labels corresponding to the outliers as the input of the cubic spline interpolation function to obtain the corrected values, and using the corrected values ​​as new data to replace the missing outliers to construct a new feature column;

[0023] S24, repeating steps S21 to S23 to obtain corrected feature columns, and recombining the corrected feature columns to obtain a preprocessed data set.

[0024] Furthermore, in step S3, the steps for implementing the CEEMDAN frequency division algorithm to divide the light intensity into multiple modules are as follows:

[0025] S31, add different white noise sequences to the original light intensity x(t) to form multiple noise-enhanced signals; let the added noise be ω i (t), we get:

[0026] x i (t)=x(t)+ω i (t)

[0027] Among them, x i (t) is the i-th noise-enhanced signal;

[0028] S32, performing EMD decomposition on each noise-enhanced signal to obtain an intrinsic mode function (IMF);

[0029] S33, average the first IMFs obtained from all noise-enhanced signals to obtain the first IMF1 of the original light intensity:

[0030]

[0031] Where M is the number of noise-enhanced signals;

[0032] S34, subtract the first IMF1 from the original light intensity x(t) to obtain the residual signal:

[0033] r i (t) = x(t) - IMF1

[0034] Among them, r i (t) represents the residual signal;

[0035] S35, for the remaining signal r i (t) Repeat steps S31 to S34, continue adding white noise and performing EMD decomposition to obtain the IMF of the residual signal until all IMFs are extracted and the residual signal becomes a monotonic function or white noise;

[0036] S36, synthesize all IMFs to obtain the complete original light intensity decomposition result.

[0037] Furthermore, in step S3, for the A-BiGRU point prediction model, the meteorological data X is used as input and the light intensity decomposition state is used as output y i ; Input the point training set of each module into the A-BiGRU point prediction model for training, use the point test set of each module for point prediction, evaluate the A-BiGRU point prediction model, and obtain the final A-BiGRU point prediction model corresponding to each module.

[0038] Furthermore, in step S4, the detailed steps for obtaining the target light intensity prediction interval are as follows:

[0039] S41: The pre-processed historical meteorological data X is input into the A-BiGRU point prediction model to obtain the light intensity point prediction value Y. i ;

[0040] S42, the light intensity point prediction result Y i The error value from the actual light intensity Y is recorded as output δ. A new interval prediction dataset is constructed using the historical meteorological data X and the error value δ. The interval prediction dataset is divided into an interval training set and an interval test set.

[0041] S43, using historical meteorological data X as input and error value δ as output, inputting the interval training set into the NGBoost interval prediction model, training the NGBoost interval prediction model, and performing interval prediction using the interval test set to obtain a trained NGBoost interval prediction model;

[0042] S44, inputting the preprocessed historical meteorological data X and the point prediction error δ into the trained NGBoost interval prediction model to obtain the probability distribution parameters mean μ and standard deviation σ required to construct the target light intensity prediction interval;

[0043] S45, obtain the target light intensity prediction interval based on the probability distribution parameter mean μ and standard deviation σ:

[0044] I α =[μ-z α2 σ,μ+z α2 σ]

[0045] Among them, α is the confidence interval, I α is the target prediction interval, z α2 is the corresponding standard score.

[0046] Compared with the prior art, the present invention has the following significant effects:

[0047] 1. The present invention detects outliers in the data using the isolation forest method and then corrects them using the cubic spline interpolation method, effectively avoiding the impact of errors in the original data set on the prediction results;

[0048] 2. The CEEMDAN frequency division algorithm is used to divide the dataset into multiple modules according to light intensity, and the A-BiGRU point prediction model is used to predict each module, making the prediction of light intensity more targeted and flexible, while also improving the versatility of the model;

[0049] 3. Input the obtained point prediction results into the NGBoost interval prediction model to obtain the interval prediction results of light intensity, and then obtain the distribution of light intensity, providing a measure of uncertainty and enhancing the interpretability of the prediction results;

[0050] 4. The MPPT algorithm is used to predict the light intensity range and obtain the upper and lower limits of the photovoltaic cell's power generation range. The appropriate energy storage battery capacity is selected, which reduces the cost requirements of the energy storage system and reduces costs.

[0051] 5. The logic of energy storage control is set. When the power generation is greater than the predicted power generation, the energy storage system is charged; otherwise, the energy storage system is discharged, so that the final output power of the system is close to the predicted value, thereby improving the accuracy of the prediction and meeting the grid connection requirements. BRIEF DESCRIPTION OF THE DRAWINGS

[0052] Figure 1 is a flow chart of the present invention;

[0053] Figure 2 Schematic diagram of photovoltaic battery-energy storage control system;

[0054] Figure 3 This is the flow chart of the conductance increment method;

[0055] Figure 4 This is the control block diagram of the energy storage system;

[0056] Figure 5 Schematic diagram of photovoltaic cell-energy storage control simulation system;

[0057] Figure 6 This is the simulation experiment result diagram under charging mode;

[0058] Figure 7 This is a schematic diagram of the DC power input for the hardware experiment in charging mode;

[0059] Figure 8 Schematic diagram for battery input of hardware experiment in charging mode;

[0060] Figure 9 This is a schematic diagram of the hardware experimental system output in charging mode;

[0061] Figure 10 This is the result of hardware simulation experiment of discharge mode;

[0062] Figure 11 This is the schematic diagram of the DC power supply input for the discharge mode hardware experiment;

[0063] Figure 12 Schematic diagram of battery input for the discharge mode hardware experiment;

[0064] Figure 13 This is the output diagram of the discharge mode hardware experimental system;

[0065] Among them, in charging mode, CD1 is the input voltage, CD2 is the input current, and CD3 is the input power; in discharge mode, FD1 is the output voltage, FD2 is the output current, and FD3 is the output power. DETAILED DESCRIPTION

[0066] The present invention will be described in further detail below with reference to the accompanying drawings and specific implementations.

[0067] In order to solve the existing problems, the present invention proposes a method that combines artificial intelligence algorithms with energy storage hardware. Through the introduction of the energy storage hardware system, the prediction accuracy of the artificial intelligence algorithm is greatly improved while also improving the versatility of the method. For different application scenarios, the prediction method of the present invention can also be directly put into use without changing the artificial intelligence algorithm architecture and the energy storage hardware system.

[0068] The present invention first obtains a relatively accurate light intensity prediction result through a combined algorithm, and then corrects the prediction result of photovoltaic power generation by connecting the photovoltaic power generation system with the energy storage system, so as to greatly improve the accuracy of photovoltaic power generation prediction and meet the photovoltaic grid-connected accuracy requirements.

[0069] like Figure 1The flowchart of the present invention is shown. The prediction method proposed by the present invention specifically includes the following steps:

[0070] Step 1: Collect the original data set, including historical meteorological data (such as humidity, temperature, wind direction, air pressure, etc.) and light intensity data, divide the original data set into multiple groups of feature columns according to time and various characteristics (such as humidity, temperature, wind direction, air pressure, light intensity, etc.), and use the isolation forest method to detect outliers for each feature column.

[0071] The implementation steps for outlier detection for each feature column using the isolation forest method are as follows:

[0072] Step 11: put n samples of a single feature column into the root node of the tree;

[0073] Step 12: The isolation forest algorithm randomly specifies a dimension and randomly generates a cut point in the current node data. The cut point is generated between the maximum and minimum values ​​of the specified dimension in the current node data.

[0074] Step 13: Generate a hyperplane using the cutting point to divide the current node data space into two subspaces: put the data less than p in the specified dimension into the left leaf of the current node, and put the data greater than or equal to p into the right leaf of the current node;

[0075] Step 14: Repeat steps 12 and 13 in the leaf nodes, and continuously construct new leaf nodes until each leaf node has only one data, that is, it can no longer be cut or the leaf node has reached the specified height, and n isolation forests are obtained;

[0076] Step 15: Repeat steps 11 to 14 above to filter out outliers in each feature column using the isolation forest method.

[0077] In step 2, the outliers of each feature column are replaced using the cubic spline interpolation method, and then all feature columns are merged to obtain the preprocessed data set.

[0078] The steps to implement cubic spline interpolation to replace outliers in each feature column are as follows:

[0079] Step 21: For a single feature column, record the sequence label corresponding to the outlier and remove the outlier.

[0080] Step 22: construct a cubic spline interpolation function for the feature column after removing outliers, with the sequence label as input and the sequence value as output;

[0081] Step 23: Use the sequence labels corresponding to the outliers recorded in step 21 as the input of the cubic spline interpolation function to obtain the corrected values, and use the corrected values ​​as new data to replace the missing outliers and construct a new feature column;

[0082] Step 24: Repeat steps 21 to 23 to obtain the corrected feature columns, and recombine the corrected feature columns to obtain the preprocessed data set.

[0083] Step 3: After preprocessing (after isolation forest and cubic spline interpolation), the data set is decomposed into multiple modules using the CEEMDAN (Complete Dataset Empirical Mode Decomposition with Adaptive Noise) decomposition algorithm, which are denoted as module 1, module 2, ..., module n.

[0084] The steps to implement the CEEMDAN frequency division algorithm to divide the light intensity into multiple modules are as follows:

[0085] Step 31: Add white noise

[0086] Add different white noise sequences to the original light intensity x(t) to form multiple noise-enhanced signals. Assume that the added noise is ω i (t), we get:

[0087] x i (t)=x(t)+ω i (t)(1)

[0088] Among them, x i (t) is the i-th noise-enhanced signal.

[0089] Step 32: Perform EMD decomposition

[0090] For each noise-enhanced signal x i (t) Perform EMD decomposition to obtain the intrinsic mode function (IMF).

[0091] Step 33, calculate the first IMF

[0092] Average the first IMFs obtained from all noise-enhanced signals to obtain the first IMF1 of the original light intensity:

[0093]

[0094] Where M is the number of noise-enhanced signals.

[0095] Step 34, calculate the residual signal

[0096] Subtracting the first IMF1 from the original light intensity x(t) yields the residual signal:

[0097] ri (t) = x(t) - IMF1 (3)

[0098] Among them, r i (t) represents the residual signal.

[0099] Step 35, iterative decomposition

[0100] For the residual signal r i (t) Repeat step 34, continue adding white noise and performing EMD decomposition to obtain the IMF of the remaining signal until all IMFs are extracted and the remaining signal becomes a monotonic function or white noise.

[0101] Step 36: Synthesize the final IMF

[0102] All IMFs are synthesized to obtain the complete original light intensity decomposition result.

[0103] For each module after decomposition, the data of each module is divided into a point training set and a point test set, with meteorological data as input X and light intensity decomposition as output y. i , input the point training set of each module into the A-BiGRU point prediction model for training, use the point test set of each module for point prediction, evaluate the A-BiGRU point prediction model, and obtain the final A-BiGRU point prediction model corresponding to each module; then use the entire data set of each module as the input of the A-BiGRU point prediction model to obtain the point prediction results of each module, and then linearly add the point prediction results of each module to obtain the final point prediction light intensity result Y i .

[0104] Step 4: Use the NGBoost interval prediction model to perform interval prediction on the light intensity to obtain the interval predicted light intensity result.

[0105] There is an inevitable error between the point prediction result and the actual light intensity. Taking the historical meteorological data X as input, the point prediction result Y i The error value with the actual light intensity Y is used as the output δ, and the output δ is divided into an interval training set and an interval test set. The interval training set is input into the NGBoot interval prediction model for training, and then the interval test set is used for interval prediction. The NGBoot interval prediction model is evaluated to obtain the final trained NGBoost interval prediction model. Further calculation can obtain the lower limit value low, upper limit value high of the light intensity prediction interval and the final prediction value pre of the light intensity.

[0106] The detailed steps to obtain the target light intensity prediction interval are as follows:

[0107] Step 41: The historical meteorological data X of the pre-processed (isolation forest and cubic spline interpolation) dataset is input into the trained A-BiGRU point prediction model to obtain the light intensity point prediction value Y corresponding to the dataset. i ;

[0108] Step 42, the point prediction result Y obtained in step 41 i The error value from the actual light intensity Y is recorded as output δ. A new interval prediction dataset is constructed using the historical meteorological data X and the error value δ for interval prediction. The interval prediction dataset is divided into an interval training set and an interval test set.

[0109] Step 43: Using the historical meteorological data X as input and the error value δ as output, the interval training set is input into the NGBoost interval prediction model, the NGBoost interval prediction model is trained, the interval test set is used for interval prediction, and the NGBoost interval prediction model is evaluated to obtain the final trained NGBoost interval prediction model.

[0110] Step 44: Input the historical meteorological data X and the point prediction error δ of the preprocessed (isolation forest and cubic spline interpolation) data set into the trained NGBoost interval prediction model to obtain the probability distribution parameters mean μ and standard deviation σ required to construct the target light intensity prediction interval;

[0111] Step 45: Obtain the target light intensity prediction interval based on the probability distribution parameter mean μ and standard deviation σ:

[0112] I α =[μ-z α2 σ,μ+z α2 σ](4)

[0113] Among them, α is the confidence interval, I α is the target prediction interval, z α2 is the corresponding standard score.

[0114] The lower limit of the light intensity prediction interval is low, that is, Y i +μ-z α2 σ, upper limit high is Y i +μ+z α2 σ, the final predicted value of light intensity pre, that is, Y i +μ.

[0115] At this time, there is still an inevitable error between the predicted light intensity value pre and the actual light intensity value real, and there are even a few time points with large errors. Therefore, the photovoltaic-energy storage system is subsequently used to correct the results.

[0116] Step 5: Convert the light intensity into photovoltaic power generation through calculation, and use the MPPT algorithm to obtain the predicted photovoltaic power generation and the photovoltaic power generation corresponding to the upper and lower limits of the light intensity range.

[0117] Photovoltaic power generation is directly affected by light intensity, and can be roughly considered a proportional relationship. In this embodiment, the conductance increment method within the MPPT algorithm is used to convert light intensity into photovoltaic power generation. The photovoltaic power generation lower limit (Low), upper limit (High), predicted value (Pre), and actual value (Real) are derived from the light intensity lower limit (Low), upper limit (High), predicted value (Pre), and actual value (Real).

[0118] Conductivity increment method as attached Figure 3 As shown, according to the PU curve of photovoltaic cells, in order to maximize the power generation, it is necessary to keep the power generation at the maximum power point, that is, when dP / dU = 0. Taking the derivative of the formula P = UI, we get dP / dU = I + U * dI / dU

[0119] 1) When dP / dU>0, then dI / dU>-I / U, the energy storage system operates on the left side of the maximum power point. At this time, the reference voltage U ref (k+1)=U ref (k)+dU ref

[0120] 2) When dP / dU<0, then dI / dU<-I / U, and the energy storage system works to the right of the maximum power point. At this time, the reference voltage U ref (k+1)=U ref (k)-dU ref

[0121] 3) When dP / dU=0, then dI / dU=-I / U, and the energy storage system works at the maximum power point. At this time, the reference voltage U ref (k+1)=U ref (k)

[0122] Through the above operations, the photovoltaic power generation can be controlled at the maximum power value, and the photovoltaic power generation lower limit value Low, upper limit value High, predicted value Pre, and real value Real can be obtained.

[0123] Step 6: Select appropriate energy storage capacity based on the upper and lower limits of photovoltaic power generation and build an energy storage control prediction system; set the energy storage control logic to charge the energy storage when the actual power generation is greater than the predicted power generation; otherwise, discharge the energy storage.

[0124] Select the appropriate energy storage size based on the maximum value of the difference between the lower limit value Low and the upper limit value High of the photovoltaic power generation, and select capacitors (C1, C2, C3, C4), inductors (L1, L2) and switches (S1, S2, S3, S4) to build the following Figure 2 The photovoltaic-energy storage system shown in the figure has a bidirectional DC / DC converter control strategy at the photovoltaic cell end using the conductance increment method. The control strategy at the energy storage system end is to charge the battery when Real is greater than Pre; conversely, the battery discharges when Real is less than or equal to Pre. The control block diagram of the bidirectional DC / DC converter is shown in the attached figure. Figure 4 As shown in the figure, the control method adopts voltage and current dual closed-loop PI control, and its control process is:

[0125] The photovoltaic cell output voltage of the Pre is set as the voltage reference value, and the bus terminal output voltage is collected. The difference between the two is PI-adjusted to obtain the battery output current reference value. The difference between the two and the collected battery output current is then PI-adjusted to obtain the switch tube duty cycle. Then, the control signal of the third switch tube S3 is obtained after PWM wave adjustment, and the control signal of the fourth switch tube S4 is obtained after inversion, so as to change the bus terminal output voltage and the battery output current until the bus terminal output voltage converges with the reference voltage.

[0126] Through the above operations, when the photovoltaic cell output power Real, the final output power at the bus terminal can converge with the photovoltaic power generation Pre, so as to achieve the goal of significantly improving the prediction accuracy.

[0127] Attachment Figure 5 For the system model built by simulink simulation, Figure 6 A schematic diagram of the simulation results of an example where Pre is less than Real obtained through artificial intelligence (Real is 52.21W, Pre is 41.37W) and Figure 10 A schematic diagram of the simulation results of an example where Pre is greater than Real obtained through artificial intelligence (Real is 48.49W, Pre is 68.96W); Figure 7 、 Figure 8 、 Figure 9 They are respectively the DC power input diagram of the hardware experiment where Pre is smaller than Real instance, the battery input diagram of the hardware experiment, and the system output diagram of the hardware experiment. Figure 11 、 Figure 12 、 Figure 13 They are respectively the hardware experiment DC power input diagram, hardware experiment battery input diagram and hardware experiment system output diagram of the Pre greater than Real instance. Figure 6 、 Figure 7 、 Figure 8 、 Figure 9It can be seen that when the Pre obtained by artificial intelligence is smaller than Real, the proposed method can make the final output power of the system close to Pre, which greatly improves the prediction accuracy. Figure 10 、 Figure 11 、 Figure 12 、 Figure 13 It can be seen that when the Pre obtained through artificial intelligence is greater than the Real, the prediction method of the present invention can also make the system's final output power close to the Pre, thereby significantly improving the prediction accuracy. From the simulation and experimental results, it can be seen that the prediction method of the present invention can effectively and significantly improve the accuracy of photovoltaic power generation prediction by combining artificial intelligence algorithms with energy storage system control.

Claims

1. A photovoltaic power generation prediction method based on NGBoost algorithm and energy storage control, characterized in that: The steps are as follows: S1, the collected original data set is divided into multiple groups of feature columns according to time and features, and the isolation forest method is used to detect outliers for each feature column; S2, uses the cubic spline interpolation method to replace the outliers in each feature column, and then merges all feature columns to obtain the preprocessed data set; S3, using the CEEMDAN frequency division algorithm, decomposes the preprocessed light intensity into multiple modules; then divides the data of each module into a point training set and a point test set, trains and evaluates the A-BiGRU point prediction model respectively, and saves the best A-BiGRU point prediction model; uses meteorological data X as the input of the A-BiGRU point prediction model, linearly adds the point prediction results of each module, and obtains the point prediction light intensity result Y i ; S4, using the preprocessed data set and the point predicted light intensity result, performs interval prediction on the predicted light intensity through the NGBoost interval prediction model to obtain the interval predicted light intensity result; S5, using the MPPT algorithm to convert the light intensity into photovoltaic power generation, and obtain the predicted photovoltaic power generation and the photovoltaic power generation corresponding to the upper and lower limits of the light intensity range; S6, select the appropriate energy storage capacity based on the upper and lower limits of power generation, and build an energy storage control prediction system; set the energy storage control logic as follows: when the actual power generation is greater than the predicted power generation, the energy storage is charged; when the actual power generation is less than or equal to the predicted power generation, the energy storage is discharged; In step S3, the implementation steps of using the CEEMDAN frequency division algorithm to divide the light intensity into multiple modules are as follows: S31, add different white noise sequences to the original light intensity x(t) to form multiple noise-enhanced signals; let the added noise be ω i (t), we get: x i (t)=x(t)+ω i (t) Among them, x i (t) is the i-th noise-enhanced signal; S32, performing EMD decomposition on each noise-enhanced signal to obtain an intrinsic mode function (IMF); S33, average the first IMFs obtained from all noise-enhanced signals to obtain the first IMF1 of the original light intensity: Where M is the number of noise-enhanced signals; S34, subtract the first IMF1 from the original light intensity x(t) to obtain the residual signal: r i (t)=x(t)-IMF1 Among them, r i (t) represents the residual signal; S35, for the remaining signal r i (t) Repeat steps S31 to S34, continue adding white noise and performing EMD decomposition to obtain the IMF of the residual signal until all IMFs are extracted and the residual signal becomes a monotonic function or white noise; S36, synthesize all IMFs to obtain the complete original light intensity decomposition result; In step S3, for the A-BiGRU point prediction model, the meteorological data X is used as input and the light intensity decomposition state is used as output y i ; Input the point training set of each module into the A-BiGRU point prediction model for training, use the point test set of each module for point prediction, evaluate the A-BiGRU point prediction model, and obtain the final A-BiGRU point prediction model corresponding to each module; In step S4, the detailed steps for obtaining the target light intensity prediction interval are as follows: S41: The pre-processed historical meteorological data X is input into the A-BiGRU point prediction model to obtain the light intensity point prediction value Y. i ; S42, the light intensity point prediction result Y i The error value from the actual light intensity Y is recorded as output δ. A new interval prediction dataset is constructed using the historical meteorological data X and the error value δ. The interval prediction dataset is divided into an interval training set and an interval test set. S43, using historical meteorological data X as input and error value δ as output, inputting the interval training set into the NGBoost interval prediction model, training the NGBoost interval prediction model, and performing interval prediction using the interval test set to obtain a trained NGBoost interval prediction model; S44, inputting the preprocessed historical meteorological data X and the point prediction error δ into the trained NGBoost interval prediction model to obtain the probability distribution parameters mean μ and standard deviation σ required to construct the target light intensity prediction interval; S45, obtain the target light intensity prediction interval based on the probability distribution parameter mean μ and standard deviation σ: I α =[μ-z α2 s,m+z α2 [s] Among them, α is the confidence interval, I α is the target prediction interval, z α2 is the corresponding standard score.

2. The photovoltaic power generation prediction method based on the NGBoost algorithm and energy storage control according to claim 1 is characterized in that: The original data set includes historical meteorological data and light intensity data, and the features include humidity, temperature, wind direction, and air pressure. The steps for implementing outlier detection for each feature column using the isolation forest method are as follows: S11, put n samples of a single feature column into the root node of the tree; S12, the isolation forest algorithm randomly specifies a dimension and randomly generates a cut point in the current node data. The cut point is generated between the maximum and minimum values ​​of the specified dimension in the current node data; S13: Generate a hyperplane based on the cutting point and divide the data space of the current node into two subspaces: put the data less than p in the specified dimension into the left leaf of the current node, and put the data greater than or equal to p into the right leaf of the current node; S14, looping steps S12 and S13 in the leaf nodes, continuously constructing new leaf nodes until each leaf node contains only one data, and obtaining n isolation forests; S15, repeat steps S11-S14, and filter out the outliers of each feature column using the isolation forest method.

3. The photovoltaic power generation prediction method based on the NGBoost algorithm and energy storage control according to claim 1 is characterized in that: In step S2, the implementation steps of using the cubic spline interpolation method to replace the outliers in each feature column are as follows: S21, for a single feature column, record the sequence label corresponding to the outlier and then remove the outlier; S22, constructing a cubic spline interpolation function for the feature column after removing outliers; S23, using the sequence labels corresponding to the outliers as the input of the cubic spline interpolation function to obtain the corrected values, and using the corrected values ​​as new data to replace the missing outliers to construct a new feature column; S24, repeating steps S21 to S23 to obtain corrected feature columns, and recombining the corrected feature columns to obtain a preprocessed data set.

Images