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 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.

Background

The new energy is connected into the power grid in large capacity and large scale, various comprehensive new energy power systems such as wind, light, hydrogen and the like are accelerating to form, the power grid dispatching faces new challenges, and the risk of stable operation is gradually increased.

In order to ensure safe and stable and economic operation of a large power grid, the power grid dispatching system is provided with a load prediction module, the load prediction module can make a pre-judgment before the system has large load fluctuation, and the power quality problems such as voltage sag and even interruption of the power grid are prevented by adjusting emergency control modes such as system output and load shedding, so that a solid foundation is laid for reasonable formulation of a dispatching plan.

However, unlike the flexibility and controllability of the conventional thermal power generating unit, the new energy has strong volatility and randomness, and the access of the large-scale new energy power station reduces the reliability of the scheduling while increasing the diversity of the system scheduling, how to reasonably arrange the scheduling plan on the basis of fully calling various energy forms? How do the grid operate safely and stably and economically? These become the scheduling problem of the new energy power system, and the addition of a new energy unit power prediction module is one of the keys for solving the scheduling problem.

The existing new energy prediction system is generally subjected to function customization according to power grid operation, is various in variety, and has the precision and the robustness to be improved; meanwhile, most of existing new energy prediction systems are offline models, historical data is adopted for model training, and real-time operation information such as real-time meteorological data, output adjustment strategies, sensor positions and channel information is not fully utilized, so that the model generalization capability is low, and prediction results with higher precision cannot be given according to the actual conditions of a local power grid. Based on the background, the existing dispatching platform is in an offline management mode for new energy, and power dispatching personnel cannot timely master the real-time running state of a new energy device and a system and cannot reasonably regulate and control according to current actual running output. Along with the deep digital transformation and refined operation requirements of the power grid, in order to further eliminate new energy, the control capability of regulating personnel on the operation safety risk of the new energy power grid is further improved, the operation state of the new energy power station is required to be predicted more accurately at the main station side, and the power grid dispatching plan is reasonably adjusted according to the current actual operation condition.

Disclosure of Invention

The invention aims to overcome the defects of the prior art, provides a new energy power system dispatching method based on digital twinning, which has higher prediction area coverage rate and smaller average width ratio of a prediction interval, enables a prediction model to have real-time updating property based on all levels of real-time data information existing on a dispatching side, improves the applicability and generalization of the prediction model to the situation of a local power grid, and realizes optimization and improvement of the existing dispatching system based on a new energy interval real-time prediction result.

The purpose of the invention is realized in the following way: the new energy power system dispatching method based on digital twin comprises the following steps:

s1, constructing a short-term direct probability prediction model suitable for photovoltaic power;

s2, 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;

and step S3, based on a new energy interval prediction result, a robust scheduling model is established, and the automatic power generation control (AGC) response of the unit is considered to cope with the fluctuation of the new energy output, so that the power balance of the power grid is ensured.

The step S1 of constructing a short-term direct probability prediction model suitable for photovoltaic power comprises the following steps: preprocessing an initial data set, removing abnormal values of the data set based on the base detection, and extracting strongly-correlated weather variables by gray correlation analysis; improving the NGBoost meta-model by means of a generalized natural gradient calculation method; fusion is carried out by using a Blending architecture, so that the model learning effect is further enhanced;

model dataset D contains n _D Samples, m features, i.e. d= { (x) _i ，y _i )}(x _i ∈R ^m ，y _i E R), where x _i Characterizing the eigenvector of the ith sample, y _i Representing the label value (true value) corresponding to the ith sample, i E (1, n) _D )。

Preprocessing an initial data set, removing abnormal values of the data set based on drawing base detection, and extracting strongly-correlated weather variables by gray correlation analysis comprises the following steps:

1) Normalizing the time series of each variable, and taking the kth of the n meteorological variable series as a comparison series S ^k (t) the photovoltaic power sequence is the reference sequence S ⁰ (t) calculating the difference between the two as the absolute value sequence delta ^k (t), where k ε (1, n);

Δ ^k (t)＝|S ^k (t)-S ⁰ (t)| (1)

2) Calculating the correlation coefficient eta ^k (t)：

Wherein: min (·) and Max (·) represent minimum and maximum values of the sequences; ρ is the resolution factor;

3) Solving the association degree gamma ^k ：

Wherein: t (T) _n Is the sequence length;

4) Setting a threshold value

A variable with a degree of association greater than a threshold value is selected,a new model dataset is composed.

The improvement of the NGBoost metamodel by means of a generalized natural gradient calculation method comprises:

the solution process of the natural gradient is improved, and a relation is established between a general gradient and the natural gradient through Fisher information quantity, and the method is concretely as follows:

in y _i Establishes a scoring function S (theta, y) based on the shannon information amount _i )，

S(θ,y _i )＝-log P _θ (y _i ) (4)

Wherein: p (P) _θ (y _i ) Is y _i Probability values in the predictive probability distribution; θ is a parameter vector of the predictive probability distribution;

taylor expansion is performed and the third order and the remainder are removed:

wherein: d' is theta edge

Moving an infinitely small step vector; />

Representing a natural gradient;

converting European space into statistical manifold, processing formula (5) under Riemann space,

wherein the calculation of the primary term can be simplified to:

the remainder is denoted as:

wherein:

thereby achieving the calculation of natural gradients by general gradients:

an improved NGBoost metamodel is built based on equation (10): taking theta DEG as an initial parameter vector, performing calculation until the mth iteration, and calculating y from a common gradient by a formula (10) _i And corresponding parameter vector thereof

Natural gradient +.>

And generating a new set of basis learners along the natural gradient direction, thereby realizing parameter vector update, and the final prediction result can be expressed as a formula (11):

wherein: alpha ^m Is a scale factor; beta is the unified learning rate; b (B) ^m For the unified representation of the base learners, the value of each sample point satisfies the Gaussian distribution, namely, theta= (mu, sigma), and the m training stage of theta correspondingly generates two base learners

It is uniformly expressed as +.>

/>

The Blending architecture is utilized to blend, and the further strengthening of the model learning effect comprises the following steps:

1) Raw dataset segmentation

Dividing an original training set into a sub training set DT and a test set DA according to a proportion, and defining an original prediction data set as DP;

2) Model fusion

Given the confidence level, building V NGBoost meta-models MO ₁ 、MO ₂ 、…、MO _V Learning DT by using the metamodels, and outputting a predicted result DA_ P, DP _P of DA and DP on the metamodels after training is completed; wherein DA_ P, DP _P is an initial statistical parameter vector of the predicted value corresponding to DA and DP;

the predicted mean value determined by DA_P and the actual result DA_OUT corresponding to the original DA data form a new data set, and a new meta-model MO is established _DA Training and obtaining predicted output MO _DA P; wherein MO is _DA P is the modified predicted statistical parameter vector, and MO is the same as DA_P _DA P has higher accuracy and less sharpness performance, and shows the advantage of model fusion.

MO is prepared from _DA P and DP P form new data set, and new meta-model MO is built _P Training is performed so as to output a final prediction statistical parameter vector, the upper limit and the lower limit of a predicted value under a given confidence level can be calculated through the vector, and the points are connected into a predicted value upper limit curve and a predicted value lower limit curve.

Step S2 is to construct a power digital twin system, and the decision making of assisting the power grid operation management regulation comprises the following steps of: based on the existing data information of each level of the dispatching side, the prediction model is trained and updated in real time by combining a quasi-real-time dispatching log, an overhaul list, line conditions, unit power information and the like in an EMS (energy management system) of a dispatching master station.

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

proposing the coverage rate I of the prediction area _F Average width ratio of prediction interval I _P As basic index, build the comprehensive score I _C As final index, the above index is calculated specificallyThe method comprises the following steps:

1) Prediction area coverage

By introducing I _F The accuracy of the probabilistic predictions is measured to quantify the reliability of the model, the index referencing the number of actual values falling within the confidence interval at a given confidence level, the larger the value, the more accurate the model,

wherein: n (N) _t To predict the number of samples; omega shape _i For the ith sample, whether the ith sample falls into the marked value of the confidence interval or not, the format is Boolean constant, the sample falls into 1 and does not fall into 0;

2) Average width ratio of prediction interval

By introducing I _P Measuring sharpness of probability prediction result, avoiding simple pursuit of I _F The confidence interval is too wide, the predicted result loses the reference value, I _P The larger the value is, the wider the confidence interval is, the higher the sharpness of the prediction distribution is, and the worse the prediction effect is;

wherein: i _P0 The confidence interval width under the initial parameters; u (U) _i 、L _i The upper limit and the lower limit of the confidence interval corresponding to the ith prediction sample;

3) Comprehensive score

Introduction of I _C Pair I _F I _P Performing comprehensive evaluation, I _C Higher values indicate better overall model performance, reduced sharpness while ensuring accuracy,

the construction of the power digital twin system comprises the following steps:

collecting offline information: the related primary equipment parameters of the power grid comprise plant stations, generator sets, main transformers, buses, circuits, information of safety control devices and the like; collecting quasi-real-time information: scheduling instructions, maintenance orders, operation tickets and other information; collecting real-time information: real-time meteorological data, fault discrimination of a safety control device and related line power flow, unit power and related switch state information related to a unit to be predicted.

The invention has the beneficial effects that: the invention discloses a new energy power system dispatching method based on digital twin, which comprises the following steps of S1, constructing a short-term direct probability prediction model applicable to photovoltaic power; s2, 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; step S3, a robust scheduling model is established based on a new energy interval prediction result, and automatic power generation control (AGC) response of a unit is considered to cope with new energy output fluctuation, so that power balance of a power grid is ensured; the new energy power system dispatching method based on digital twin has higher prediction area coverage rate and smaller average width ratio of the prediction interval, enables the prediction model to have real-time updating property based on the existing real-time data information of each level of the dispatching side, improves the applicability and generalization of the prediction model to the local power grid situation, and realizes the optimization and improvement of the existing dispatching system based on the real-time prediction result of the new energy interval.

Drawings

Fig. 1 is a schematic flow chart of a new energy power system dispatching method based on digital twin.

FIG. 2 is a schematic diagram of a Blending fusion step.

Detailed Description

The invention is further described below with reference to the accompanying drawings.

The new energy power system dispatching method based on digital twin is shown in the attached figure 1, and comprises the following steps:

s1, constructing a short-term direct probability prediction model suitable for photovoltaic power;

s2, 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;

and step S3, based on a new energy interval prediction result, a robust scheduling model is established, and the automatic power generation control (AGC) response of the unit is considered to cope with the fluctuation of the new energy output, so that the power balance of the power grid is ensured.

Aiming at the application defect of the ensemble learning algorithm in the probability prediction problem, the university of Steady team of AndrewY.Ng leading rank proposes a natural gradient lifting (Natural Gradient Boosting, NGboost) model, and although the popularization and application of Boosting type algorithm are realized, the following defects still exist when the actual engineering problem of photovoltaic power short-term direct probability prediction is solved: 1) The model lacks a data preprocessing link, and the generalization capability and the robustness of the NGBoost model to different photovoltaic fields are weak; 2) The natural gradient has complex calculation principle and difficult practical engineering application; 3) It is difficult for a single NGBoost meta-model to guarantee the accuracy and sharpness of probability prediction.

For better effect, the step S1 of constructing a short-term direct probability prediction model applicable to photovoltaic power includes: preprocessing an initial data set, removing abnormal values of the data set based on the base detection, and extracting strongly-correlated weather variables by gray correlation analysis; improving the NGBoost meta-model by means of a generalized natural gradient calculation method; fusion is carried out by using a Blending architecture, so that the model learning effect is further enhanced;

model dataset D contains n _D Samples, m features, i.e. d= { (x) _i ，y _i )}(x _i ∈R ^m ，y _i E R), where x _i Characterizing the eigenvector of the ith sample, y _i Representing the label value (true value) corresponding to the ith sample, i E (1, n) _D )。

Considering actual engineering conditions such as measurement errors, the initial data set has more abnormal values, which can cause the overall deviation of the prediction model. Therefore, the present item first eliminates outliers in the initial data using a box graph proposed by the collectist john's graph base.

Photovoltaic power is related to meteorological variables such as humidity and cloud cover. However, the correlation between various meteorological variables and photovoltaic power varies from one photovoltaic field to another, subject to factors such as the geographic location of the photovoltaic field and the local microclimate of the location.

For better effect, the preprocessing the initial data set, eliminating the abnormal value of the data set based on the graph base detection, and extracting the strongly-correlated weather variable by using gray correlation analysis comprises the following steps:

1) Normalizing the time series of each variable, and taking the kth of the n meteorological variable series as a comparison series S ^k (t) the photovoltaic power sequence is the reference sequence S ⁰ (t) calculating the difference between the two as the absolute value sequence delta ^k (t), where k ε (1, n);

Δ ^k (t)＝|S ^k (t)-S ⁰ (t)| (1)

2) Calculating the correlation coefficient eta ^k (t)：

Wherein: min (·) and Max (·) represent minimum and maximum values of the sequences; ρ is the resolution factor;

3) Solving the association degree gamma ^k ：

Wherein: t (T) _n Is the sequence length;

4) Setting a threshold value

And selecting variables with the association degree larger than a threshold value to form a new model data set.

The key point of NGBoost is the solution of natural gradient, however, the related concept is taken from extremely complex information geometry, which brings inconvenience to popularization and application in actual engineering.

For better effect, the improvement of the NGBoost metamodel by means of a generalized natural gradient calculation method comprises:

the solution process of the natural gradient is improved, and a relation is established between a general gradient and the natural gradient through Fisher information quantity, and the method is concretely as follows:

in y _i Establishes a scoring function S (theta, y) based on the shannon information amount _i )，

S(θ,y _i )＝-log P _θ (y _i ) (4)

Wherein: p (P) _θ (y _i ) Is y _i Probability values in the predictive probability distribution; θ is a parameter vector of the predictive probability distribution;

taylor expansion is performed and the third order and the remainder are removed:

wherein: d' is theta edge

Moving an infinitely small step vector; />

Representing a natural gradient;

converting European space into statistical manifold, processing formula (5) under Riemann space,

wherein the calculation of the primary term can be simplified to:

the remainder is denoted as:

wherein:

thereby achieving the calculation of natural gradients by general gradients:

an improved NGBoost metamodel is built based on equation (10): taking theta DEG as an initial parameter vector, performing calculation until the mth iteration, and calculating y from a common gradient by a formula (10) _i And corresponding parameter vector thereof

Natural gradient +.>

And generating a new set of basis learners along the natural gradient direction, thereby realizing parameter vector update, and the final prediction result can be expressed as a formula (11):

wherein: alpha ^m Is a scale factor; beta is the unified learning rate; b (B) ^m For the unified representation of the base learners, the value of each sample point satisfies the Gaussian distribution, namely, theta= (mu, sigma), and the m training stage of theta correspondingly generates two base learners

It is uniformly expressed as +.>

/>

The fusion of the meta-model can strengthen the learning effect, and can not cause excessive redundancy of the whole model, and is widely applied in solving the prediction problem in recent years, particularly the fusion of the Stacking model. However, stacking model fusion is too complex, and the problem of data crossing of training data referencing global statistics occurs in the training process, which is not suitable for solving the problem of probability prediction.

In order to achieve a better effect, the Blending architecture is utilized to blend, and the further strengthening of the model learning effect comprises the following steps:

1) Raw dataset segmentation

Dividing an original training set into a sub training set DT and a test set DA according to a proportion, and defining an original prediction data set as DP;

2) Model fusion

Given the confidence level, building V NGBoost meta-models MO ₁ 、MO ₂ 、…、MO _V Learning DT by using the metamodels, and outputting a predicted result DA_ P, DP _P of DA and DP on the metamodels after training is completed; wherein DA_ P, DP _P is an initial statistical parameter vector of the predicted value corresponding to DA and DP;

the predicted mean value determined by DA_P and the actual result DA_OUT corresponding to the original DA data form a new data set, and a new meta-model MO is established _DA Training and obtaining predicted output MO _DA P; wherein MO is _DA P is the modified predicted statistical parameter vector, and MO is the same as DA_P _DA P has higher accuracy and less sharpness performance, and shows the advantage of model fusion.

MO is prepared from _DA P and DP P form new data set, and new meta-model MO is built _P Training is performed so as to output a final prediction statistical parameter vector, the upper limit and the lower limit of a predicted value under a given confidence level can be calculated through the vector, and the points are connected into a predicted value upper limit curve and a predicted value lower limit curve.

In order to achieve a better effect, the step S2 is to construct a digital twin system of electric power, and the decision making for assisting the regulation and control of the power grid operation and management by using the ubiquitous internet of things and the electric power data stream through real-time situation awareness and real-time virtual deduction comprises the following steps: based on the existing data information of each level of the dispatching side, the prediction model is trained and updated in real time by combining a quasi-real-time dispatching log, an overhaul list, line conditions, unit power information and the like in an EMS (energy management system) of a dispatching master station.

In order to achieve a better effect, the step S3 of establishing a robust scheduling model based on the new energy interval prediction result, considering automatic power generation control (AGC) response of the unit to cope with new energy output fluctuation, and ensuring power balance of the power grid includes:

proposing the coverage rate I of the prediction area _F Average width ratio of prediction interval I _P As basic index, build the comprehensive score I _C As a final index, the specific calculation method of the above index is as follows:

1) Prediction area coverage

By introducing I _F The accuracy of the probabilistic predictions is measured to quantify the reliability of the model, the index referencing the number of actual values falling within the confidence interval at a given confidence level, the larger the value, the more accurate the model,

wherein: n (N) _t To predict the number of samples; omega shape _i For the ith sample, whether the ith sample falls into the marked value of the confidence interval or not, the format is Boolean constant, the sample falls into 1 and does not fall into 0;

2) Average width ratio of prediction interval

By introducing I _P Measuring sharpness of probability prediction result, avoiding simple pursuit of I _F The confidence interval is too wide, the predicted result loses the reference value, I _P The larger the value is, the wider the confidence interval is, the higher the sharpness of the prediction distribution is, and the worse the prediction effect is;

wherein: i _P0 The confidence interval width under the initial parameters; u (U) _i 、L _i The upper limit and the lower limit of the confidence interval corresponding to the ith prediction sample;

3) Comprehensive score

Introduction of I _C Pair I _F I _P Performing comprehensive evaluation, I _C Higher values indicate better overall model performance, reduced sharpness while ensuring accuracy,

for better effect, the construction of the electric power digital twin system comprises:

collecting offline information: the related primary equipment parameters of the power grid comprise plant stations, generator sets, main transformers, buses, circuits, information of safety control devices and the like; collecting quasi-real-time information: scheduling instructions, maintenance orders, operation tickets and other information; collecting real-time information: real-time meteorological data, fault discrimination of a safety control device and related line power flow, unit power and related switch state information related to a unit to be predicted.

In summary, the digital twin-based new energy power system scheduling method comprises the steps of S1, constructing a short-term direct probability prediction model applicable to photovoltaic power; s2, 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; step S3, a robust scheduling model is established based on a new energy interval prediction result, and automatic power generation control (AGC) response of a unit is considered to cope with new energy output fluctuation, so that power balance of a power grid is ensured; the new energy power system dispatching method based on digital twin has higher prediction area coverage rate and smaller average width ratio of the prediction interval, enables the prediction model to have real-time updating property based on the existing real-time data information of each level of the dispatching side, improves the applicability and generalization of the prediction model to the local power grid situation, and realizes the optimization and improvement of the existing dispatching system based on the real-time prediction result of the new energy interval.

Claims (8)

1. The new energy power system dispatching method based on digital twin is characterized by comprising the following steps of:

s1, constructing a short-term direct probability prediction model suitable for photovoltaic power;

s2, 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;

and step S3, based on a new energy interval prediction result, a robust scheduling model is established, and the automatic power generation control (AGC) response of the unit is considered to cope with the fluctuation of the new energy output, so that the power balance of the power grid is ensured.

2. The method for dispatching a new energy power system based on digital twin according to claim 1, wherein the step S1 of constructing a short-term direct probability prediction model applicable to photovoltaic power comprises: preprocessing an initial data set, removing abnormal values of the data set based on the base detection, and extracting strongly-correlated weather variables by gray correlation analysis; improving the NGBoost meta-model by means of a generalized natural gradient calculation method; fusion is carried out by using a Blending architecture, so that the model learning effect is further enhanced;

model dataset D contains n _D Samples, m features, i.e. d= { (x) _i ，y _i )}(x _i ∈R ^m ，y _i E R), where x _i Characterizing the eigenvector of the ith sample, y _i Representing the label value (true value) corresponding to the ith sample, i E (1, n) _D )。

3. The method for dispatching a new energy power system based on digital twinning according to claim 2, wherein the preprocessing the initial dataset, removing the outliers of the dataset based on the graph base detection, and extracting the strongly correlated meteorological variables by using gray correlation analysis comprises:

1) Normalizing the time series of each variable, and taking the kth of the n meteorological variable series as a comparison series S ^k (t) the photovoltaic power sequence is the reference sequence S ⁰ (t) calculating the difference between the two as the absolute value sequence delta ^k (t), where k ε (1, n);

Δ ^k (t)＝|S ^k (t)-S ⁰ (t)| (1)

2) Calculating the correlation coefficient eta ^k (t)：

Wherein: min (·) and Max (·) represent minimum and maximum values of the sequences; ρ is the resolution factor;

3) Solving the association degree gamma ^k ：

Wherein: t (T) _n Is the sequence length;

4) Setting a threshold value

And selecting variables with the association degree larger than a threshold value to form a new model data set.

4. The method for dispatching a new energy power system based on digital twinning as claimed in claim 2, wherein said improving the NGBoost metamodel by means of generalized natural gradient calculation method comprises:

the solution process of the natural gradient is improved, and a relation is established between a general gradient and the natural gradient through Fisher information quantity, and the method is concretely as follows:

in y _i Establishes a scoring function S (theta, y) based on the shannon information amount _i )，

S(θ,y _i )＝-log P _θ (y _i ) (4)

Wherein: p (P) _θ (y _i ) Is y _i Probability values in the predictive probability distribution; θ is a parameter vector of the predictive probability distribution;

taylor expansion is performed and the third order and the remainder are removed:

wherein: d' is theta edge

Moving an infinitely small step vector; />

Representing a natural gradient;

converting European space into statistical manifold, processing formula (5) under Riemann space,

wherein the calculation of the primary term can be simplified to:

the remainder is denoted as:

wherein:

thereby achieving the calculation of natural gradients by general gradients:

an improved NGBoost metamodel is built based on equation (10): taking theta DEG as an initial parameter vector, performing calculation until the mth iteration, and calculating y from a common gradient by a formula (10) _i And corresponding parameter vector thereof

Natural gradient +.>

And generating a new set of basis learners along the natural gradient direction, thereby realizing parameter vector update, and the final prediction result can be expressed as a formula (11):

wherein: alpha ^m Is a scale factor; beta is the unified learning rate; b (B) ^m For the unified representation of the base learners, the value of each sample point satisfies the Gaussian distribution, namely, theta= (mu, sigma), and the m training stage of theta correspondingly generates two base learners

It is uniformly expressed as +.>

5. The method for dispatching the new energy power system based on digital twin according to claim 2, wherein the fusing is performed by using a Blending architecture, and further strengthening the model learning effect comprises:

1) Raw dataset segmentation

Dividing an original training set into a sub training set DT and a test set DA according to a proportion, and defining an original prediction data set as DP;

2) Model fusion

Given the confidence level, building V NGBoost meta-models MO ₁ 、MO ₂ 、…、MO _V Learning DT by using the metamodels, and outputting a predicted result DA_ P, DP _P of DA and DP on the metamodels after training is completed; wherein DA_ P, DP _P is an initial statistical parameter vector of the predicted value corresponding to DA and DP;

the predicted mean value determined by DA_P and the actual result DA_OUT corresponding to the original DA data form a new data set, and a new meta-model MO is established _DA Training and obtaining predicted output MO _DA P; wherein MO is _DA P is the modified predicted statistical parameter vector, and MO is the same as DA_P _DA P has higher accuracy and less sharpness performance, and shows the advantage of model fusion.

MO is prepared from _DA P and DP P form new data set, and new meta-model MO is built _P Training is performed so as to output a final prediction statistical parameter vector, the upper limit and the lower limit of a predicted value under a given confidence level can be calculated through the vector, and the points are connected into a predicted value upper limit curve and a predicted value lower limit curve.

6. The method for dispatching the new energy power system based on digital twinning as claimed in claim 1, wherein the step S2 is to construct a power digital twinning system, and the decision making of assisting the grid operation management regulation by using ubiquitous internet of things, power data flow, real-time situation awareness and real-time virtual deduction comprises the following steps: based on the existing data information of each level of the dispatching side, the prediction model is trained and updated in real time by combining a quasi-real-time dispatching log, an overhaul list, line conditions, unit power information and the like in an EMS (energy management system) of a dispatching master station.

7. The method for dispatching the new energy power system based on digital twin according to claim 1, wherein the step S3 of establishing a robust dispatching model based on the new energy interval prediction result, considering automatic power generation control (AGC) response of the unit to cope with new energy output fluctuation, and ensuring power balance of the power grid comprises:

proposing the coverage rate I of the prediction area _F Average width ratio of prediction interval I _P As basic index, build the comprehensive score I _C As a final index, the specific calculation method of the above index is as follows:

1) Prediction area coverage

By introducing I _F The accuracy of the probabilistic predictions is measured to quantify the reliability of the model, the index referencing the number of actual values falling within the confidence interval at a given confidence level, the larger the value, the more accurate the model,

wherein: n (N) _t To predict the number of samples; omega shape _i For the ith sample, whether the ith sample falls into the marked value of the confidence interval or not, the format is Boolean constant, the sample falls into 1 and does not fall into 0;

2) Average width ratio of prediction interval

By introducing I _P Measuring sharpness of probability prediction result, avoiding simple pursuit of I _F The confidence interval is too wide, the predicted result loses the reference value, I _P The larger the value is, the wider the confidence interval is, the higher the sharpness of the prediction distribution is, and the worse the prediction effect is;

wherein: i _P0 The confidence interval width under the initial parameters; u (U) _i 、L _i The upper limit and the lower limit of the confidence interval corresponding to the ith prediction sample;

3) Comprehensive score

Introduction of I _C Pair I _F I _P Performing comprehensive evaluation, I _C Higher values indicate better overall model performance, reduced sharpness while ensuring accuracy,

8. the method for dispatching a new energy power system based on digital twinning according to claim 6, wherein the constructing the power digital twinning system comprises:

collecting offline information: the related primary equipment parameters of the power grid comprise plant stations, generator sets, main transformers, buses, circuits, information of safety control devices and the like; collecting quasi-real-time information: scheduling instructions, maintenance orders, operation tickets and other information; collecting real-time information: real-time meteorological data, fault discrimination of a safety control device and related line power flow, unit power and related switch state information related to a unit to be predicted.

Images