CN116231749A - New energy power system dispatching method based on digital twin - Google Patents

Abstract

The invention belongs to the technical field of new energy power system dispatching, and particularly relates to a new energy power system dispatching method based on digital twinning; it comprises the following steps: constructing a short-term direct probability prediction model suitable for photovoltaic power; constructing an electric power digital twin system, and utilizing ubiquitous Internet of things and electric power data flow to assist decision making of power grid operation management regulation through real-time situation awareness and real-time virtual deduction; based on the new energy interval prediction result, a robust scheduling model is established, the automatic power generation control response of the unit is considered to cope with the new energy output fluctuation, and the power balance of the power grid is ensured; the invention has higher prediction area coverage rate and smaller average width ratio of the prediction interval, based on the existing real-time data information of each level at the dispatching side, the prediction model has real-time updating property, the applicability and generalization of the prediction model to the local power grid situation are improved, and the optimization and improvement of the existing dispatching system are realized based on the real-time prediction result of the new energy interval.

Description

New energy power system dispatching method based on digital twin

Technical Field

The present invention belongs to the technical field of new energy power system dispatching, and specifically relates to a new energy power system dispatching method based on digital twins.

Background Art

New energy is being connected to the power grid on a large scale with large capacity, and a variety of integrated new energy power systems such as wind, solar, and hydrogen are being formed at an accelerated pace. Grid dispatching is facing new challenges, and the risks of stable operation are gradually increasing.

In order to ensure the safety, stability and economical operation of large power grids, power grid dispatching systems are equipped with load forecasting modules, which can make predictions before large load fluctuations occur in the system. By adjusting system output, shedding loads and other emergency control methods, it can prevent power quality problems such as voltage drops or even interruptions in the power grid, laying a solid foundation for the rational formulation of dispatching plans.

However, unlike the flexibility and controllability of traditional thermal power units, new energy has strong volatility and randomness. The access of large-scale new energy power stations increases the diversity of system dispatching while also reducing the reliability of dispatching. At this time, how to reasonably arrange the dispatching plan based on the full use of various energy forms? How to balance the safe and stable operation of the power grid and economic efficiency? These have become the difficulties in the dispatching of new energy power systems, and adding a power prediction module for new energy units is one of the keys to solving this problem.

The existing new energy prediction systems are generally customized according to the operation of the power grid, with various types, and their accuracy and robustness need to be improved; at the same time, most of the existing new energy prediction systems are offline models, which use historical data for model training and do not make full use of real-time meteorological data, output adjustment strategies, sensor locations, channel information and other real-time operation information, which leads to low model generalization ability and cannot give more accurate prediction results according to the actual situation of the local power grid. Based on the above background, the existing dispatching platform is in an offline management mode for new energy, and the power dispatching personnel cannot grasp the real-time operation status of new energy devices and systems in time, nor can they reasonably regulate according to the current actual operation output. With the deepening of the digital transformation of the power grid and the requirements for refined operation, in order to further absorb new energy and further improve the control personnel's ability to control the safety risks of the new energy power grid operation, the master station side needs to more accurately predict the operation status of new energy power stations and reasonably adjust the power grid dispatching plan according to the current actual operation.

Summary of the invention

The purpose of the present invention is to overcome the shortcomings of the prior art and to provide a new energy power system scheduling method based on digital twins, which has a higher prediction area coverage and a smaller average prediction interval width ratio. Based on the existing real-time data information at all levels on the scheduling side, the prediction model has real-time updating capabilities, thereby improving the applicability and generalization of the prediction model to local power grid conditions. Based on the real-time prediction results of new energy intervals, the existing scheduling system is optimized and improved.

The object of the present invention is achieved by: a new energy power system dispatching method based on digital twin, which comprises the following steps:

Step S1, constructing a short-term direct probability prediction model suitable for photovoltaic power;

Step S2: construct a power digital twin system, use ubiquitous Internet of Things, power data stream, and assist in decision-making for power grid operation and management through real-time situation awareness and real-time virtual simulation;

Step S3: Based on the prediction results of the new energy interval, a robust scheduling model is established, which takes into account the automatic generation control (AGC) response of the unit to cope with the fluctuation of new energy output and ensure the power balance of the power grid.

The step S1 of constructing a short-term direct probability prediction model suitable for photovoltaic power includes: preprocessing the initial data set, removing outliers from the data set based on graph-based detection, and extracting strongly correlated meteorological variables using grey correlation analysis; improving the NGBoost metamodel with the help of a generalized natural gradient calculation method; and using a blending architecture for fusion to further enhance the model learning effect;

The model dataset D ^contains _nD samples and m features, that is, D = {( _xi , _yi )}( _xi∈Rm , _yi∈R ), _wherexi represents the feature vector of the ith sample, _yi represents the label value (true value) corresponding to the ith sample, i∈(1, _nD ).

The preprocessing of the initial data set, removing outliers from the data set based on graph-based detection, and extracting strongly correlated meteorological variables using grey correlation analysis include:

1) Normalize the time series of each variable, take the kth of the n meteorological variable sequences as the comparison sequence S ^k (t), and the photovoltaic power sequence as the reference sequence S ⁰ (t), and calculate the difference between the two as the absolute value sequence Δ ^k (t), where k∈(1,n);

Δ ^k (t)＝|S ^k (t)-S ⁰ (t)| (1)

2) Calculate the correlation coefficient η ^k (t):

Where: Min(·) and Max(·) represent the minimum and maximum values of the sequence; ρ is the resolution coefficient;

3) Solve the correlation degree γ ^k :

Where: _Tn is the sequence length;

选取关联度大于阈值的变量，组成新的模型数据集。4) Set the threshold

Variables with correlation greater than the threshold are selected to form a new model data set.

The improved NGBoost metamodel by using the generalized natural gradient calculation method includes:

The solution process of the natural gradient is improved, and the connection between the general gradient and the natural gradient is established through the Fisher information, as follows:

The scoring function S(θ, y _i ) is established based on the Shannon information of y _i .

S(θ,y _i )＝-log P _θ (y _i ) (4)

Where: P _θ (y _i ) is the probability value of y _i in the predicted probability distribution; θ is the parameter vector of the predicted probability distribution;

Perform Taylor expansion and discard the remainder of order 3 and above:

移动的无穷小步长向量；

表示自然梯度；Where: d' is the

infinitesimal step vector of movement;

represents the natural gradient;

Transform the Euclidean space into a statistical manifold and process equation (5) in Riemann space:

The calculation of the first-order term can be simplified as:

The remainder is expressed as:

Where:

This allows the calculation of the natural gradient through the general gradient:

的自然梯度

并沿该自然梯度方向生成一组新的基学习器，从而实现参数向量更新，最终预测结果可以表示为式(11)：Based on formula (10), the improved NGBoost metamodel is established: Taking θ° as the initial parameter vector, the calculation is performed to the mth iteration, and y _i and its corresponding parameter vector are calculated by ordinary gradient through formula (10)

Natural gradient

A new set of base learners is generated along the natural gradient direction to update the parameter vector. The final prediction result can be expressed as formula (11):

将其统一表示为

Where: ^αm is the scaling factor; β is the uniform learning rate; ^Bm is the uniform representation of the base learner, and the value of each sample point satisfies the Gaussian distribution, that is, θ = (μ, σ), and the mth training stage of θ corresponds to the generation of two base learners

It is uniformly expressed as

The use of the Blending architecture for fusion further enhances the model learning effect, including:

1) Original dataset segmentation

The original training set is divided into a sub-training set DT and a test set DA in proportion, and the original prediction data set is defined as DP;

2) Model Fusion

Given a confidence level, construct V NGBoost meta-models MO ₁ , MO ₂ , …, MO _V , use these meta-models to learn DT, and after training, output the prediction results DA_P and DP_P of DA and DP on the meta-model; where DA_P and DP_P are the initial statistical parameter vectors of the corresponding prediction values of DA and DP;

The predicted mean determined by DA_P and the actual result DA_OUT corresponding to the original DA data are combined into a new data set, and a new meta-model MO _DA is established for training to obtain the predicted output MO _DA _P; among them, MO _DA _P is the corrected prediction statistical parameter vector. Compared with DA_P, MO _DA _P has higher accuracy and smaller sharpness, reflecting the advantages of model fusion.

MO _DA _P and DP_P are combined into a new data set, and a new meta-model MO _P is established for training, so as to output the final prediction statistical parameter vector. Through this vector, the upper and lower limits of the prediction value under a given confidence level can be calculated, and these points are connected to form the upper and lower limit curves of the prediction value.

The step S2 constructs a power digital twin system, and utilizes the ubiquitous Internet of Things and power data streams to assist in decision-making for power grid operation, management and control through real-time situation awareness and real-time virtual deduction, including: based on the existing data information at all levels on the dispatching side, combined with quasi-real-time dispatching logs, maintenance orders, and line conditions and unit power information in the dispatching master station EMS system, real-time training and updating of the prediction model.

The step S3 establishes a robust dispatch model based on the prediction result of the new energy interval, considers the automatic generation control (AGC) response of the unit to cope with the fluctuation of the new energy output, and ensures the power balance of the power grid, including:

The prediction area coverage rate I _F and the prediction interval average width ratio I _P are proposed as basic indicators, and the comprehensive score I _C is established as the final indicator. The specific calculation methods of the above indicators are as follows:

1) Prediction of regional coverage

By introducing _IF to measure the accuracy of probability prediction results, the reliability of the model can be quantified. This indicator uses the number of actual values falling within the confidence interval at a given confidence level as a reference. The larger the value, the more accurate the model.

Where: _Nt is the number of predicted samples; _Ωi is the mark value of whether the i-th sample falls into the confidence interval, the format is a Boolean constant, the sample falls into the confidence interval is counted as 1, and the sample does not fall into the confidence interval is counted as 0;

2) Average width of the prediction interval

By introducing _IP to measure the sharpness of probability prediction results, we can avoid the situation where the confidence interval is too wide due to the simple pursuit of _IF , and the prediction results lose their reference value. The larger the _IP value, the wider the confidence interval, the sharper the prediction distribution, and the worse the prediction effect.

Where: I _P0 is the confidence interval width under the initial parameters; U _i , L _i are the upper and lower limits of the confidence interval corresponding to the i-th prediction sample;

3) Comprehensive score

_IC is introduced to comprehensively evaluate I _F and I _P. The higher the _IC value, the better the overall performance of the model. It reduces the sharpness while ensuring the accuracy.

The construction of the power digital twin system includes:

Collect offline information: parameters of primary equipment in the power grid, including plant, generator set, main transformer, bus, line, safety and control device information, etc.; collect quasi-real-time information: dispatch instructions, maintenance orders, operation tickets and other information; collect real-time information: real-time meteorological data, safety and control device fault diagnosis, and related line flow, unit power, and related switch status information involved in the predicted unit.

Beneficial effects of the present invention: The digital twin-based new energy power system scheduling method of the present invention comprises step S1, constructing a short-term direct probability prediction model suitable for photovoltaic power; step S2, constructing a power digital twin system, utilizing ubiquitous Internet of Things and power data stream, through real-time situation perception and real-time virtual deduction, to assist in decision-making for power grid operation and management; step S3, based on the new energy interval prediction results, establishing a robust scheduling model, considering the automatic generation control (AGC) response of the unit to cope with the new energy output fluctuation, and ensuring the power balance of the power grid; the digital twin-based new energy power system scheduling method of the present invention has a higher prediction area coverage rate and a smaller average prediction interval width ratio, and based on the existing real-time data information of various levels on the scheduling side, the prediction model has real-time updateability, improves the applicability and generalization of the prediction model to the local power grid conditions, and realizes the optimization and improvement of the existing scheduling system based on the real-time prediction results of the new energy interval.

BRIEF DESCRIPTION OF THE DRAWINGS

FIG1 is a flow chart of a new energy power system dispatching method based on digital twins according to the present invention.

FIG. 2 is a schematic diagram of the blending steps.

DETAILED DESCRIPTION

The present invention will be further described below in conjunction with the accompanying drawings.

The new energy power system dispatching method based on digital twin, as shown in Figure 1, includes the following steps:

Step S1, constructing a short-term direct probability prediction model suitable for photovoltaic power;

Step S2: construct a power digital twin system, use ubiquitous Internet of Things, power data stream, and assist in decision-making for power grid operation and management through real-time situation awareness and real-time virtual simulation;

Step S3: Based on the prediction results of the new energy interval, a robust scheduling model is established, which takes into account the automatic generation control (AGC) response of the unit to cope with the fluctuation of new energy output and ensure the power balance of the power grid.

In response to the application defects of ensemble learning algorithms in probability prediction problems, a Stanford University team led by Andrew Y. Ng proposed a Natural Gradient Boosting (NGBoost) model. Although it has achieved the promotion and application of Boosting algorithms, it still has the following defects when solving the actual engineering problems of short-term direct probability prediction of photovoltaic power: 1) The model lacks data preprocessing, and the generalization ability and robustness of the NGBoost model for different photovoltaic fields are weak; 2) The calculation principle of the natural gradient itself is complex, and it is difficult to apply it in actual engineering; 3) A single NGBoost meta-model is difficult to guarantee the accuracy and sharpness of probability prediction.

For better results, the step S1 constructs a short-term direct probability prediction model suitable for photovoltaic power, including: preprocessing the initial data set, removing data set outliers based on graph-based detection, extracting strongly correlated meteorological variables using grey correlation analysis; improving the NGBoost metamodel with the help of a generalized natural gradient calculation method; and using the Blending architecture for fusion to further enhance the model learning effect;

The model dataset D ^contains _nD samples and m features, that is, D = {( _xi , _yi )}( _xi∈Rm , _yi∈R ), _wherexi represents the feature vector of the ith sample, _yi represents the label value (true value) corresponding to the ith sample, i∈(1, _nD ).

Considering actual engineering situations such as measurement errors, there are many outliers in the initial data set, which will cause the overall deviation of the prediction model. Therefore, this project first uses the box plot proposed by statistician John Tukey to eliminate outliers in the initial data.

Photovoltaic power is related to meteorological variables such as humidity and cloud cover. However, due to factors such as the geographical location of the photovoltaic field and the local microclimate of the location, the correlation between various meteorological variables and photovoltaic power varies among different photovoltaic fields.

For better results, the initial data set is preprocessed, outliers in the data set are removed based on graph-based detection, and strongly correlated meteorological variables are extracted using grey correlation analysis, including:

1) Normalize the time series of each variable, take the kth of the n meteorological variable sequences as the comparison sequence S ^k (t), and the photovoltaic power sequence as the reference sequence S ⁰ (t), and calculate the difference between the two as the absolute value sequence Δ ^k (t), where k∈(1,n);

Δ ^k (t)＝|S ^k (t)-S ⁰ (t)| (1)

2) Calculate the correlation coefficient η ^k (t):

Where: Min(·) and Max(·) represent the minimum and maximum values of the sequence; ρ is the resolution coefficient;

3) Solve the correlation degree γ ^k :

Where: _Tn is the sequence length;

选取关联度大于阈值的变量，组成新的模型数据集。4) Set the threshold

Variables with correlation greater than the threshold are selected to form a new model data set.

The key to NGBoost lies in the solution of natural gradient. However, the relevant concepts are derived from extremely complex information geometry, which brings inconvenience to its promotion and application in practical engineering.

For better results, the improved NGBoost metamodel by using the generalized natural gradient calculation method includes:

The solution process of the natural gradient is improved, and the connection between the general gradient and the natural gradient is established through the Fisher information, as follows:

The scoring function S(θ, y _i ) is established based on the Shannon information of y _i .

S(θ,y _i )＝-log P _θ (y _i ) (4)

Where: P _θ (y _i ) is the probability value of y _i in the predicted probability distribution; θ is the parameter vector of the predicted probability distribution;

Perform Taylor expansion and discard the remainder of order 3 and above:

移动的无穷小步长向量；

表示自然梯度；Where: d' is the

infinitesimal step vector of movement;

represents the natural gradient;

Transform the Euclidean space into a statistical manifold and process equation (5) in Riemann space:

The calculation of the first-order term can be simplified as:

The remainder is expressed as:

Where:

This allows the calculation of the natural gradient through the general gradient:

的自然梯度

并沿该自然梯度方向生成一组新的基学习器，从而实现参数向量更新，最终预测结果可以表示为式(11)：Based on formula (10), the improved NGBoost metamodel is established: Taking θ° as the initial parameter vector, the calculation is performed to the mth iteration, and y _i and its corresponding parameter vector are calculated by ordinary gradient through formula (10)

Natural gradient

A new set of base learners is generated along the natural gradient direction to update the parameter vector. The final prediction result can be expressed as formula (11):

将其统一表示为

Where: ^αm is the scaling factor; β is the uniform learning rate; ^Bm is the uniform representation of the base learner, and the value of each sample point satisfies the Gaussian distribution, that is, θ = (μ, σ), and the mth training stage of θ corresponds to the generation of two base learners

It is uniformly expressed as

Fusion of meta-models can enhance learning effects without causing excessive redundancy in the overall model. In recent years, it has been widely used in solving prediction problems, especially Stacking model fusion. However, Stacking model fusion is too complex, and data crossing problems will occur during training when the training data references global statistics, which is not suitable for solving probability prediction problems.

In order to achieve better results, the Blending architecture is used to further enhance the model learning effect, including:

1) Original dataset segmentation

The original training set is divided into a sub-training set DT and a test set DA in proportion, and the original prediction data set is defined as DP;

2) Model Fusion

Given a confidence level, construct V NGBoost meta-models MO ₁ , MO ₂ , …, MO _V , use these meta-models to learn DT, and after training, output the prediction results DA_P and DP_P of DA and DP on the meta-model; where DA_P and DP_P are the initial statistical parameter vectors of the corresponding prediction values of DA and DP;

The predicted mean determined by DA_P and the actual result DA_OUT corresponding to the original DA data are combined into a new data set, and a new meta-model MO _DA is established for training to obtain the predicted output MO _DA _P; among them, MO _DA _P is the corrected prediction statistical parameter vector. Compared with DA_P, MO _DA _P has higher accuracy and smaller sharpness, reflecting the advantages of model fusion.

MO _DA _P and DP_P are combined into a new data set, and a new meta-model MO _P is established for training, so as to output the final prediction statistical parameter vector. Through this vector, the upper and lower limits of the prediction value under a given confidence level can be calculated, and these points are connected to form the upper and lower limit curves of the prediction value.

For better results, step S2 constructs a power digital twin system, utilizes the ubiquitous Internet of Things and power data streams, and assists in decision-making for power grid operation, management and control through real-time situation awareness and real-time virtual deduction, including: based on the existing data information at all levels on the dispatching side, combined with quasi-real-time dispatching logs, maintenance orders, and line conditions and unit power information in the dispatching master station EMS system, real-time training and updating of the prediction model.

For better results, step S3 establishes a robust dispatch model based on the prediction results of the new energy interval, considers the automatic generation control (AGC) response of the unit to cope with the fluctuation of new energy output, and ensures the power balance of the power grid, including:

The prediction area coverage rate I _F and the prediction interval average width ratio I _P are proposed as basic indicators, and the comprehensive score I _C is established as the final indicator. The specific calculation methods of the above indicators are as follows:

1) Prediction of regional coverage

By introducing _IF to measure the accuracy of probability prediction results, the reliability of the model can be quantified. This indicator uses the number of actual values falling within the confidence interval at a given confidence level as a reference. The larger the value, the more accurate the model.

Where: _Nt is the number of predicted samples; _Ωi is the mark value of whether the i-th sample falls into the confidence interval, the format is a Boolean constant, the sample falls into the confidence interval is counted as 1, and the sample does not fall into the confidence interval is counted as 0;

2) Average width of the prediction interval

By introducing _IP to measure the sharpness of probability prediction results, we can avoid the situation where the confidence interval is too wide due to the simple pursuit of _IF , and the prediction results lose their reference value. The larger the _IP value, the wider the confidence interval, the sharper the prediction distribution, and the worse the prediction effect.

Where: I _P0 is the confidence interval width under the initial parameters; U _i , L _i are the upper and lower limits of the confidence interval corresponding to the i-th prediction sample;

3) Comprehensive score

_IC is introduced to comprehensively evaluate I _F and I _P. The higher the _IC value, the better the overall performance of the model. It reduces the sharpness while ensuring the accuracy.

For better results, the construction of the power digital twin system includes:

Collect offline information: parameters of primary equipment in the power grid, including plant, generator set, main transformer, bus, line, safety and control device information, etc.; collect quasi-real-time information: dispatch instructions, maintenance orders, operation tickets and other information; collect real-time information: real-time meteorological data, safety and control device fault diagnosis, and related line flow, unit power, and related switch status information involved in the predicted unit.

In summary, the new energy power system scheduling method based on digital twins of the present invention includes step S1, constructing a short-term direct probability prediction model suitable for photovoltaic power; step S2, constructing a power digital twin system, using ubiquitous Internet of Things, power data flow, through real-time situation perception and real-time virtual deduction, to assist in the decision-making of power grid operation and control; step S3, based on the new energy interval prediction results, establishing a robust scheduling model, considering the automatic generation control (AGC) response of the unit to cope with the new energy output fluctuations, to ensure the power balance of the power grid; the new energy power system scheduling method based on digital twins of the present invention has a higher prediction area coverage rate and a smaller prediction interval average width ratio. Based on the existing real-time data information of various levels on the scheduling side, the prediction model has real-time updateability, improves the applicability and generalization of the prediction model to the local power grid conditions, and based on the real-time prediction results of the new energy interval, realizes the optimization and improvement of the existing scheduling system.

Claims (8)

1. A new energy power system dispatching method based on digital twins, characterized in that it includes the following steps: Step S1, constructing a short-term direct probability prediction model suitable for photovoltaic power; Step S2: construct a power digital twin system, use ubiquitous Internet of Things, power data stream, and assist in decision-making for power grid operation and management through real-time situation awareness and real-time virtual simulation; Step S3: Based on the prediction results of the new energy interval, a robust scheduling model is established, which takes into account the automatic generation control (AGC) response of the unit to cope with the fluctuation of new energy output and ensure the power balance of the power grid. 2. The method for dispatching a new energy power system based on digital twins according to claim 1 is characterized in that the step S1 of constructing a short-term direct probability prediction model suitable for photovoltaic power comprises: preprocessing the initial data set, removing outliers from the data set based on graph-based detection, and extracting strongly correlated meteorological variables using grey correlation analysis; improving the NGBoost metamodel with the help of a generalized natural gradient calculation method; and using a blending architecture for fusion to further enhance the model learning effect; The model dataset D ^contains _nD samples and m features, that is, D = {( _xi , _yi )}( _xi∈Rm , _yi∈R ), _wherexi represents the feature vector of the ith sample, _yi represents the label value (true value) corresponding to the ith sample, i∈(1, _nD ). 3. The method for dispatching a new energy power system based on digital twins according to claim 2 is characterized in that the preprocessing of the initial data set, removing abnormal values of the data set based on graph-based detection, and extracting strongly correlated meteorological variables using grey correlation analysis include: 1) Normalize the time series of each variable, take the kth of the n meteorological variable sequences as the comparison sequence S ^k (t), and the photovoltaic power sequence as the reference sequence S ⁰ (t), and calculate the difference between the two as the absolute value sequence Δ ^k (t), where k∈(1,n); Δ ^k (t)＝|S ^k (t)-S ⁰ (t)| (1) 2) Calculate the correlation coefficient η ^k (t):

Where: Min(·) and Max(·) represent the minimum and maximum values of the sequence; ρ is the resolution coefficient; 3) Solve the correlation degree γ ^k :

Where: _Tn is the sequence length; 4) Set the threshold

Variables with correlation greater than the threshold are selected to form a new model data set. 4. The method for dispatching a new energy power system based on digital twins according to claim 2, wherein the step of improving the NGBoost metamodel by means of a generalized natural gradient calculation method comprises: The solution process of the natural gradient is improved, and the connection between the general gradient and the natural gradient is established through the Fisher information, as follows: The scoring function S(θ, y _i ) is established based on the Shannon information of y _i . S(θ,y _i )＝-log P _θ (y _i ) (4) Where: P _θ (y _i ) is the probability value of y _i in the predicted probability distribution; θ is the parameter vector of the predicted probability distribution; Perform Taylor expansion and discard the remainder of order 3 and above:

Where: d′ is the

infinitesimal step vector of movement;

represents the natural gradient; Transform the Euclidean space into a statistical manifold and process equation (5) in Riemann space:

The calculation of the first-order term can be simplified as:

The remainder is expressed as:

Where:

This allows the calculation of the natural gradient through the general gradient:

Based on formula (10), the improved NGBoost metamodel is established: Taking θ° as the initial parameter vector, the calculation is performed to the mth iteration, and y _i and its corresponding parameter vector are calculated by ordinary gradient through formula (10)

Natural gradient

A new set of base learners is generated along the natural gradient direction to update the parameter vector. The final prediction result can be expressed as formula (11):

Where: ^αm is the scaling factor; β is the uniform learning rate; ^Bm is the uniform representation of the base learner, and the value of each sample point satisfies the Gaussian distribution, that is, θ = (μ, σ), and the mth training stage of θ corresponds to the generation of two base learners

It is uniformly expressed as

5. The method for dispatching a new energy power system based on digital twins according to claim 2, characterized in that the fusion using the Blending architecture to further enhance the model learning effect includes: 1) Original dataset segmentation The original training set is divided into a sub-training set DT and a test set DA in proportion, and the original prediction data set is defined as DP; 2) Model Fusion Given a confidence level, construct V NGBoost meta-models MO ₁ , MO ₂ , …, MO _V , use these meta-models to learn DT, and after training, output the prediction results DA_P and DP_P of DA and DP on the meta-model; where DA_P and DP_P are the initial statistical parameter vectors of the corresponding prediction values of DA and DP; The predicted mean determined by DA_P and the actual result DA_OUT corresponding to the original DA data are combined into a new data set, and a new meta-model MO _DA is established for training to obtain the predicted output MO _DA _P; among them, MO _DA _P is the corrected prediction statistical parameter vector. Compared with DA_P, MO _DA _P has higher accuracy and smaller sharpness, reflecting the advantages of model fusion. MO _DA _P and DP_P are combined into a new data set, and a new meta-model MO _P is established for training, so as to output the final prediction statistical parameter vector. Through this vector, the upper and lower limits of the prediction value under a given confidence level can be calculated, and these points are connected to form the upper and lower limit curves of the prediction value. 6. The new energy power system dispatching method based on digital twins as described in claim 1 is characterized in that the step S2 constructs a power digital twin system, utilizes the ubiquitous Internet of Things, power data flow, and assists in the decision-making of power grid operation and management through real-time situation awareness and real-time virtual deduction, including: based on the existing various levels of data information on the dispatching side, combined with quasi-real-time dispatching logs, maintenance orders, and line conditions and unit power information in the dispatching master station EMS system, real-time training and updating of the prediction model. 7. The method for dispatching a new energy power system based on digital twins according to claim 1 is characterized in that, in step S3, a robust dispatching model is established based on the prediction results of the new energy interval, and the automatic generation control (AGC) response of the unit is considered to cope with the fluctuation of the new energy output, and ensuring the power balance of the power grid includes: The prediction area coverage rate I _F and the prediction interval average width ratio I _P are proposed as basic indicators, and the comprehensive score I _C is established as the final indicator. The specific calculation methods of the above indicators are as follows: 1) Prediction of regional coverage By introducing _IF to measure the accuracy of probability prediction results, the reliability of the model can be quantified. This indicator uses the number of actual values falling within the confidence interval at a given confidence level as a reference. The larger the value, the more accurate the model.

Where: _Nt is the number of predicted samples; _Ωi is the mark value of whether the i-th sample falls into the confidence interval, the format is a Boolean constant, the sample falls into the confidence interval is counted as 1, and the sample does not fall into the confidence interval is counted as 0; 2) Average width of the prediction interval By introducing _IP to measure the sharpness of probability prediction results, we can avoid the situation where the confidence interval is too wide due to the simple pursuit of _IF , and the prediction results lose their reference value. The larger the _IP value, the wider the confidence interval, the sharper the prediction distribution, and the worse the prediction effect.

Where: I _P0 is the confidence interval width under the initial parameters; U _i , L _i are the upper and lower limits of the confidence interval corresponding to the i-th prediction sample; 3) Comprehensive score _IC is introduced to comprehensively evaluate I _F and I _P. The higher the _IC value, the better the overall performance of the model. It reduces the sharpness while ensuring the accuracy.

8. The method for dispatching a new energy power system based on digital twins according to claim 6, wherein the construction of a power digital twin system comprises: Collect offline information: parameters of primary equipment in the power grid, including plant, generator set, main transformer, bus, line, safety and control device information, etc.; collect quasi-real-time information: dispatch instructions, maintenance orders, operation tickets and other information; collect real-time information: real-time meteorological data, safety and control device fault diagnosis, and related line flow, unit power, and related switch status information involved in the predicted unit.

Images