CN109995095B - Intelligent operation control method of power distribution and utilization system based on data driving - Google Patents

Intelligent operation control method of power distribution and utilization system based on data driving Download PDF

Info

Publication number
CN109995095B
CN109995095B CN201910248085.8A CN201910248085A CN109995095B CN 109995095 B CN109995095 B CN 109995095B CN 201910248085 A CN201910248085 A CN 201910248085A CN 109995095 B CN109995095 B CN 109995095B
Authority
CN
China
Prior art keywords
value
training
data
current
voltage
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Active
Application number
CN201910248085.8A
Other languages
Chinese (zh)
Other versions
CN109995095A (en
Inventor
邓卫
裴玮
赵振兴
张学
孔力
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Institute of Electrical Engineering of CAS
Original Assignee
Institute of Electrical Engineering of CAS
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by Institute of Electrical Engineering of CAS filed Critical Institute of Electrical Engineering of CAS
Priority to CN201910248085.8A priority Critical patent/CN109995095B/en
Publication of CN109995095A publication Critical patent/CN109995095A/en
Application granted granted Critical
Publication of CN109995095B publication Critical patent/CN109995095B/en
Active legal-status Critical Current
Anticipated expiration legal-status Critical

Links

Images

Classifications

    • H—ELECTRICITY
    • H02—GENERATION; CONVERSION OR DISTRIBUTION OF ELECTRIC POWER
    • H02J—CIRCUIT ARRANGEMENTS OR SYSTEMS FOR SUPPLYING OR DISTRIBUTING ELECTRIC POWER; SYSTEMS FOR STORING ELECTRIC ENERGY
    • H02J3/00—Circuit arrangements for ac mains or ac distribution networks
    • H—ELECTRICITY
    • H02—GENERATION; CONVERSION OR DISTRIBUTION OF ELECTRIC POWER
    • H02J—CIRCUIT ARRANGEMENTS OR SYSTEMS FOR SUPPLYING OR DISTRIBUTING ELECTRIC POWER; SYSTEMS FOR STORING ELECTRIC ENERGY
    • H02J5/00—Circuit arrangements for transfer of electric power between ac networks and dc networks
    • H—ELECTRICITY
    • H02—GENERATION; CONVERSION OR DISTRIBUTION OF ELECTRIC POWER
    • H02J—CIRCUIT ARRANGEMENTS OR SYSTEMS FOR SUPPLYING OR DISTRIBUTING ELECTRIC POWER; SYSTEMS FOR STORING ELECTRIC ENERGY
    • H02J2203/00—Indexing scheme relating to details of circuit arrangements for AC mains or AC distribution networks
    • H02J2203/20—Simulating, e g planning, reliability check, modelling or computer assisted design [CAD]

Landscapes

  • Engineering & Computer Science (AREA)
  • Power Engineering (AREA)
  • Supply And Distribution Of Alternating Current (AREA)

Abstract

The invention relates to a power distribution and utilization system intelligent operation control method based on data driving, which is characterized in that a control decision is generated by an intelligent data analysis mining method of a power distribution and utilization system through automatic training of a deep belief network by utilizing the existing operation data, and a decision service is provided for improving the safety stability margin of the power distribution and utilization system; data acquisition and preprocessing are carried out, input quantity is optimized, and training speed is accelerated; unsupervised pre-training is carried out on the training set, and neural network parameters are acquired more efficiently by combining with improved self-adaptive weight learning rate; carrying out supervised fine adjustment, and optimizing the initialization weight and the threshold of the neural network by using the improved particle swarm algorithm based on the adaptive inertial weight factor updating method and the fitness value, so that the training speed of the subsequent BP algorithm is accelerated, and the prediction effect is improved; the invention greatly reduces the complex design workload of system control decision and improves the overall efficiency.

Description

Intelligent operation control method of power distribution and utilization system based on data driving
Technical Field
The invention relates to a data-driven intelligent operation control method for a power distribution and utilization system, and belongs to the technical field of power distribution and utilization system control.
Background
Distributed renewable energy power generation is widely concerned in the global scope, large-scale access demonstration including electric vehicles, distributed energy storage, combined heat and power, renewable energy and the like is gradually developed, the advantages of reducing transformation links and improving the overall operation efficiency along with direct current access are increasingly highlighted, and distributed energy such as photovoltaic and wind power are gathered in a direct current mode and flexibly accessed to form a new development trend. On the other hand, direct current equipment such as electric vehicles and LED lighting is widely used, and the strong application requirements of both the supply and demand further promote the rapid development and demonstration popularization of low-voltage direct current distribution and power utilization systems and multi-terminal alternating current and direct current distribution and power utilization systems.
A plurality of power electronic converters in the multi-terminal alternating current-direct current power distribution and utilization system, in particular to a converter (VSC) for bearing power flow control between a direct current network and an alternating current system, can perform mutual compensation on power flows among the alternating current systems through the direct current network through coordination control, and enables the whole system to have wider power sharing capability and power supply capability. The low-voltage multi-terminal direct-current system can provide functions of emergency control, power oscillation damping, dynamic voltage multi-terminal mutual support and the like, impact and influence of disturbance on the system are reduced, and overall stability is improved.
Fig. 1 illustrates a structure in which 3 AC systems are interconnected by multi-terminal dc, wherein a single AC system AC is interconnected with a dc network by a VSC, and AC systems 1, 2, 3 are connected to the dc network by VSC1, VSC2, VCS3 on a certain feeder line, respectively, wherein AC sides of VSC1, VSC2, VSC3 are respectively connected to AC feeder lines of AC systems 1, 2, 3, and at the same time, dc sides of the three are connected to a dc bus via a line of a certain length. The direct current network can integrate photovoltaic and other distributed renewable energy sources, an energy storage system, an electric automobile and a direct current load, wherein when the direct current voltage level of the equipment is not matched with the voltage level of the direct current bus, the DC/DC converter can be configured to convert, and part of alternating current equipment such as wind power can be connected to the direct current bus through the AC/DC converter.
When the VSC3 is used as a master station and adopts a constant direct-current voltage control strategy, the direct-current side voltage can be regarded as constant and is equivalent to a direct-current voltage source, the direct-current voltage source is connected with a direct-current bus through a direct-current line, imRespectively, the line current values. VSC1, VSC2 all adopt the constant power control strategy as slave stations, connect the direct current bus, U through the direct current circuit respectively0 s1,Us1,is1And Ps1Respectively representing the steady-state voltage value of the DC side of the VSC1, the value after the change of the DC side voltage, the line current value and the value of the interaction power with the AC system 1, U0 s2,Us2,is2And Ps2The dc side steady state voltage value, the dc side voltage changed value, the line current value and the interaction power value with the ac system 2 of the VSC2 are indicated, respectively. The energy storage device, the distributed power supply and the direct current load are connected into a direct current bus, wherein U0 dc,UdcRepresenting the steady state voltage value of the DC bus, the value after the voltage change of the DC bus, PdcRepresenting the energy storage device power. DC load andthe distributed power sources are aggregated and expressed by equivalent load.
On the basis of normal operation, an active control strategy is adopted, and the centralized energy management and control center can generate additional power instructions of all control objects through comprehensive analysis and complex calculation on the basis of acquiring operation information of the power distribution and utilization system, particularly current, direct current voltage and other data in a direct current network, so that the damping characteristic of the system is improved. On the basis that the existing local controllers of the control objects such as the energy storage device, the VSC1 and the VSC2 are not changed, the centralized energy management and control center generates delta Pdc,ΔPs1,ΔPs2In which Δ PdcFor additional power of energy storage devices, Δ Ps1Additional power, Δ P, for VSC1s2Additional power, Δ P, for VSC2dc,ΔPs1,ΔPs2And respectively superposed to the existing scheduling reference values of the energy storage device, the VSC1 and the VSC2 to realize the improvement of the stability margin.
However, the design of the control decision by depending on global information needs to acquire the operation information of the power distribution and utilization system in real time, the requirements on the real-time performance of communication and the rapidity of instruction calculation are high, and how to generate the additional power instruction often needs to be designed in a complex way and is strongly related to a main circuit model, line parameters and the like of the system, and the flexibility and the expandability of the method are weak due to the factors.
Disclosure of Invention
The invention solves the problems: the method overcomes the defects of the prior art, provides the intelligent operation control method of the power distribution and utilization system based on data driving, and guarantees the intelligent and safe operation of the power distribution and utilization system.
The technical scheme of the invention is as follows: a power distribution and utilization system intelligent operation control method based on data driving utilizes an intelligent data analysis mining method of a power distribution and utilization system, utilizes existing operation data to generate a control decision through automatic training of a deep belief network, and provides decision service for improving the safety and stability margin of the power distribution and utilization system; data acquisition and preprocessing are carried out, input quantity is optimized, and training speed is accelerated; unsupervised pre-training is carried out on the training set, and neural network parameters are acquired more efficiently by combining with improved self-adaptive weight learning rate; carrying out supervised fine adjustment, and optimizing the initialization weight and the threshold of the neural network by using the improved particle swarm algorithm based on the adaptive inertial weight factor updating method and the fitness value, so that the training speed of the subsequent BP algorithm is accelerated, and the prediction effect is improved; the invention greatly reduces the complex design workload of system control decision and improves the overall efficiency.
The method comprises the following specific steps:
the first step is as follows: data acquisition and pretreatment, namely acquiring data from historical operating data of the multi-terminal AC/DC power distribution system, wherein the data are a plurality of groups of TxAnd TyA plurality of groups TxAnd TyAre combined into training sets T _ set, each set TxAnd TyForming a training data T, each training data T ═ Tx,TY}，TxIs the running data of a multi-terminal AC/DC distribution power system, TYIs corresponding to TxAn additional power command for the power distribution and utilization system; wherein:
Figure GDA0002145252050000031
TY＝{ΔPdc,ΔPs1,ΔPs2}
wherein i0 m,i0 s1,i0 s2,U0 dc,U0 s1,U0 s2,Udc,Us1,Us2Respectively showing the steady state current value of a main station line, the steady state current value of a slave station VSC1 line, the steady state current value of a slave station VSC2 line, the steady state voltage value of a direct current bus, the steady state voltage value of the direct current side of VSC1, the steady state voltage value of the direct current side of VSC2, the value after the change of the direct current bus voltage, the value after the change of the direct current side voltage of VSC1, the value after the change of the direct current side voltage of VSC2, delta PdcFor additional power of energy storage devices, Δ Ps1Additional power, Δ P, for VSC1s2Additional power for VSC 2;
preprocessing the training data to obtain training neural network input values formed by corresponding training data;
the second step is that: unsupervised pre-training, training the training neural network input value correspondingly formed by each training data obtained in the first step by adopting Deep Belief Network (DBN), and completing i-learning by using a DBN deep learning structure and a self-adaptive weight learning rate0 m,i0 s1,i0 s2,U0 dc,U0 s1,U0 s2,Udc,Us1,Us2、ΔPdc、ΔPs1、ΔPs2Forming a neural network;
the third step: aiming at the neural network obtained in the second step, correcting the initialization weight and the threshold of the neural network by adopting a Particle Swarm Optimization (PSO) algorithm based on self-adaptive inertia weight factor updating and fitness value, and then carrying out supervised fine tuning by using a BP algorithm to obtain an improved neural network;
the fourth step: acquiring current operation data R of the power distribution and utilization system based on the improved neural network obtained in the third stepx，
Figure GDA0002145252050000032
Inputting the current additional power instruction into an improved neural network to obtain a current additional power instruction for intelligent operation control of the power distribution and consumption system based on data driving:
RY＝{ΔPdc,now,ΔPs1,now,ΔPs2,now}
wherein i0 m,now,i0 s1,now,i0 s2,now,U0 dc,now,U0 s1,now,U0 s2,now,Udc,now,Us1,now,Us2,nowThe steady state current value of the current state of the line of the master station, the steady state current value of the current state of the line of the slave station VSC1, the steady state current value of the current state of the line of the slave station VSC2, the steady state voltage value of the current state of the direct current bus, the steady state voltage value of the direct current side of the current state of the VSC1, the steady state voltage value of the direct current side of the current state of the VSC2, the current value of the voltage of the direct current bus, and the voltage of theThe current value, the dc side voltage current value of the VSC 2. Delta Pdc,nowAdditional power, Δ P, for the current state of the energy storage devices1,nowAdditional power, Δ P, for the current state of VSC1s2,nowIs the additional power for the current state of the VSC 2.
In the first step, data are obtained from historical operation data of the multi-terminal AC/DC power distribution system, and the data are a plurality of groups of TxAnd TyA plurality of groups TxAnd TyAre combined into training sets T _ set, each set TxAnd TyForming a training data T, each training data T ═ Tx,TY}，TxIs the running data of a multi-terminal AC/DC distribution power system, TYIs corresponding to TxAn additional power command for the power distribution and utilization system; wherein:
Figure GDA0002145252050000041
TY＝{ΔPdc,ΔPs1,ΔPs2}
for a certain training data T, if
Figure GDA0002145252050000042
Wherein U isdc,min、Udc,max、Udc,ratedThe direct current bus voltage minimum value, the direct current bus voltage maximum value and the direct current bus voltage rated value are respectively obtained; u shapes1,min、Us1,max、Us1,ratedThe minimum value of the direct-current side voltage, the maximum value of the direct-current side voltage and the rated value of the direct-current side voltage of the VSC1 are respectively; u shapes2,min、Us2,max、Us2,ratedThe dc-side voltage minimum value, the dc-side voltage maximum value, and the dc-side voltage rated value of the VSC2 are respectively, and α is a start threshold value.
Then set in the training data T:
Figure GDA0002145252050000043
Figure GDA0002145252050000044
Figure GDA0002145252050000045
Figure GDA0002145252050000046
X5＝Udc
Figure GDA0002145252050000047
X7＝Us1
Figure GDA0002145252050000048
X9＝Us2
wherein, XiAn ith input value representing a training neural network, and if not, skipping the training data T;
and processing the 1 st to the last training data in the training set T _ set according to the steps to obtain the training neural network input values correspondingly formed by the training data.
In the second step, the self-adaptive weight learning rate satisfies: when the updating direction of a certain weight value is consistent twice continuously, the step length is increased:
(m+1)＝(1-log(m))*(m)
when the updating direction of a certain weight value is inconsistent twice continuously, the step length is reduced:
(m+1)＝(1+log(m))*(m)
wherein (m +1) represents the weight learning rate at the m +1 th iteration, and (m) represents the weight learning rate at the m th iteration;
through the training process, the weights of the edges between different nodes in each hidden layer and the threshold values of each layer are obtained, and the corresponding neural network is obtained.
The improved particle swarm PSO algorithm is realized based on a self-adaptive inertia weight factor updating method and a fitness value, wherein the formula of the self-adaptive inertia weight factor updating method is as follows:
Figure GDA0002145252050000051
wherein: ω (t +1) represents the PSO particle swarm inertial weight factor at the t-th iteration, ωmaxAnd ωminThe values of the upper boundary value and the lower boundary value of the inertia weight factor are 0.4 and 0.9; mmaxIs the maximum allowable iteration number; | v (t) | is the two-norm of the velocity vector at the t-th iteration; i V (t-1) I is the two-norm of the velocity vector at the time of the t-1 iteration;
the fitness value is composed of a prediction error absolute value sum of training data, a neural network connection weight square sum and a hidden layer threshold square sum, and the calculation formula is as follows:
Figure GDA0002145252050000052
in the formula, F represents the adaptability value of the PSO algorithm, N is the number of nodes of an output layer, and k1、k2Is a scale factor, ωij(l) As a weight of an edge between the jth node of the l layer and the ith node of the l-1 layer in the hidden layer, bj(l) Is the threshold value of the jth node of the ith layer in the hidden layer.
In the fourth step, the current operation data R of the power distribution and utilization system is collected based on the improved neural network obtained in the third stepx，
Figure GDA0002145252050000053
Inputting the current additional power instruction into an improved neural network, and further acquiring a current additional power instruction for intelligent operation control of the power distribution and consumption system based on data driving:
RY＝{ΔPdc,now,ΔPs1,now,ΔPs2,now}
compared with the prior art, the invention has the advantages that:
(1) low-voltage direct current distribution and power utilization and multi-terminal alternating current and direct current distribution and power utilization systems are rapidly developed, the system structure is increasingly complex, the difficulty of fine integral modeling of the system is continuously increased, and the accuracy of system operation control is further influenced. The invention utilizes the data mining method, can utilize the existing historical operation data to reduce the dependence on the traditional physical model, and automatically generates the operation control decision through data driving, thereby obviously reducing the basic design workload of system operation control.
(2) According to the invention, through an intelligent data analysis and mining method of the power distribution and utilization system, the existing operation data is utilized to generate the control decision through the automatic training of the deep belief network, and the decision service is provided for improving the safety and stability margin of the power distribution and utilization system. Wherein: optimizing input quantity through data acquisition and preprocessing to accelerate training speed; unsupervised pre-training is carried out on the training set, and neural network parameters are acquired more efficiently by combining with improved self-adaptive weight learning rate; on the basis, supervised fine tuning is carried out, and the initialization weight and the threshold of the neural network are optimized through the improved particle swarm optimization based on the adaptive inertial weight factor updating method and the fitness value, so that the training speed of the subsequent BP algorithm is accelerated, and the prediction effect is improved. The invention only excavates the information knowledge contained in the existing operation data through a large amount of operation data and utilizes the information knowledge, thereby greatly reducing the complex design workload of system control decision and improving the overall efficiency.
(3) The intelligent data analysis and mining method for the power distribution and utilization system is provided, the defects of the traditional physical model are overcome by combining data, the decision is formed by data driving, the physical model of the power distribution and utilization system is not needed to be known, information and knowledge contained in the power distribution and utilization system are mined only through a large amount of existing operation data, the control decision is generated through automatic training of a deep belief network, the safety stability margin of the power distribution and utilization system is improved, and the intelligent safe operation of the power distribution and utilization system is guaranteed.
Drawings
FIG. 1 is a diagram of a multi-terminal AC/DC distribution system;
FIG. 2 is a flow chart of the intelligent operation control method of the present invention;
FIG. 3 is a DBN structure in the present invention;
FIG. 4 is a flow chart of the PSO algorithm for improving particle swarm optimization according to the present invention.
Detailed Description
The present invention will be described in detail below with reference to the accompanying drawings and examples.
As shown in fig. 2, the intelligent operation control method of the power distribution system based on data driving of the present invention specifically includes the following steps:
the first step is as follows: data acquisition and preprocessing
Obtaining data from historical operating data of a multi-terminal AC/DC distribution power system, wherein the data comprise a plurality of groups TxAnd TyA plurality of groups TxAnd TyAre combined into training sets T _ set, each set TxAnd TyForming a training data T, each training data T ═ Tx,TY}，TxIs the running data of a multi-terminal AC/DC distribution power system, TYIs corresponding to TxAn additional power command for the power distribution and utilization system; wherein:
Figure GDA0002145252050000061
TY＝{ΔPdc,ΔPs1,ΔPs2}
i0 m,i0 s1,i0 s2,U0 dc,U0 s1,U0 s2,Udc,Us1,Us2the steady state current value of the line of the main station is respectively shown, the steady state current value of the line of the secondary station VSC1, the steady state current value of the line of the secondary station VSC2, the steady state voltage value of a direct current bus, the steady state voltage value of the direct current side of VSC1, the steady state voltage value of the direct current side of VSC2, the value after the voltage of the direct current bus changes, the value after the voltage of the direct current side of VSC1 changes, and the value after the voltage of the direct current side of VS. Delta PdcFor additional power of energy storage devices, Δ Ps1Additional power, Δ P, for VSC1s2Is the additional power of the VSC 2.
For a certain training data T, if
Figure GDA0002145252050000071
Wherein U isdc,min、Udc,max、Udc,ratedThe direct current bus voltage minimum value, the direct current bus voltage maximum value and the direct current bus voltage rated value are respectively obtained; u shapes1,min、Us1,max、Us1,ratedThe minimum value of the direct-current side voltage, the maximum value of the direct-current side voltage and the rated value of the direct-current side voltage of the VSC1 are respectively; u shapes2,min、Us2,max、Us2,ratedThe dc-side voltage minimum value, the dc-side voltage maximum value, and the dc-side voltage rated value of the VSC2, respectively. Alpha is the activation threshold.
Then set in the training data T:
Figure GDA0002145252050000072
Figure GDA0002145252050000073
Figure GDA0002145252050000074
Figure GDA0002145252050000075
X5＝Udc
Figure GDA0002145252050000076
X7＝Us1
Figure GDA0002145252050000077
X9＝Us2
wherein, XiRepresenting the ith input value of the training neural network. If not, the training data T is skipped.
And processing the 1 st to the last training data in the training set T _ set according to the steps to obtain the training neural network input values correspondingly formed by the training data.
The second step is that: unsupervised pre-training
A Deep Belief Network (DBN) is used to train a training neural network input value correspondingly formed by each training data, the DBN uses a Restricted Boltzmann Machine (RBM) as a basic unit, and is generally formed by a CD-based RBM unsupervised pre-training algorithm and a BP-based supervised fine tuning algorithm.
Wherein, the structure of the DBN of the present invention is shown in fig. 3. The DBN is a deep neural network structure formed by stacking a plurality of RBMs, with the output of a lower layer RBM serving as the input of an upper layer RBM. Learning of label-free data can be completed through the DBN deep learning structure, and then a neural network with certain cognitive recognition capability is formed. Wherein v in the parameters of RBM is a visible unit, h is a hidden unit, WijIs v isiAnd hjThe weight of the edge in between, where viDenotes the ith visible cell, where hjDenotes the jth hidden unit, aiThreshold representing the ith visible cell, bjRepresenting the threshold of the jth hidden unit. The update rule of the Contrast Divergence (CD) weight and the threshold based on Hinton satisfies the following conditions:
ΔWij＝(<vihj>data—<vihj>recon)
wherein the rate is learned for the weight,<vihj>datain order to be expected for the distribution of data,<vihj>reconthe distribution defined for the model after one-step reconstruction.
In the RBM learning process, in order to avoid conflict caused by overlarge step length, the invention provides an improved self-adaptive weight learning rate, namely when the updating direction of a certain weight is consistent twice continuously, the step length is increased:
(m+1)＝(1-log(m))*(m)
when the updating direction of a certain weight value is inconsistent twice continuously, the step length is reduced:
(m+1)＝(1+log(m))*(m)
where (m +1) represents the weight learning rate at the m +1 th iteration, and (m) represents the weight learning rate at the m-th iteration.
Through the training process, the weights of the edges between different nodes in each hidden layer and the threshold values of each layer are obtained, and the corresponding neural network is obtained.
The third step: with supervised fine tuning
Aiming at the neural network obtained in the second step, an improved Particle Swarm Optimization (PSO) algorithm is adopted to correct the initialization weight and the threshold of the neural network, then a BP algorithm is utilized to carry out supervised fine tuning, and through correction and Optimization, the training speed is accelerated and the prediction effect of the neural network is improved.
The PSO algorithm has the basic idea that m particles are randomly generated in an S-dimensional space, each particle represents a vector formed by a group of quantities to be solved and is expressed as Xi＝(xi1，xi2，…，xis) I-1, 2, …, m, the position of each particle represents a potential solution of the system. The "jump" velocity of the ith particle is also an S-dimensional vector, denoted as Vi＝(vi1，vi2，…，vis) The optimum position of the ith particle searched so far is Pi＝(pi1，pi2，…，pis) The optimal position searched for by the whole population so far is Pg＝(pg1，pg2，…，pgn). In each iteration, each particle passes through a trace PiAnd PgThese 2 "extrema" update the spatial position and velocity:
Vi(t+1)＝ωVi(t)+c1r1(t)[Pi(t)-Xi(t)]+c2r2(t)[Pg(t)-Xi(t)]
Xi(t+1)＝Xi(t)+Vi(t+1)
wherein t represents the number of iterations, Xi(t)、Vi(t) spatial position of ith particle in the tth iterationSetting and speed; omega is an inertial weight factor, c1、c2The value range of the learning factor is [1,3 ]]，r1(t) and r2(t) is [0, 1 ]]A random number in between.
The invention adopts a self-adaptive inertia weight factor updating method which comprises the following steps:
Figure GDA0002145252050000081
wherein: ω (t +1) represents the PSO particle swarm inertial weight factor at the t-th iteration, ωmaxAnd ωminUpper and lower boundary values for the inertial weight factor, typically taken to be 0.4 and 0.9; mmaxIs the maximum allowable iteration number; | v (t) | is the two-norm of the velocity vector at the t-th iteration; and | V (t-1) | is the two-norm of the velocity vector at the t-1 th iteration.
The fitness value of the PSO algorithm provided by the invention consists of the sum of absolute prediction errors of training data, the sum of squares of connection weights of a neural network and the sum of squares of thresholds of a hidden layer, and the calculation formula is as follows:
Figure GDA0002145252050000091
in the formula, F represents the adaptability value of the PSO algorithm, N is the number of nodes of an output layer, and k1、k2Is a scale factor, ωij(l) As a weight of an edge between the jth node of the l layer and the ith node of the l-1 layer in the hidden layer, bj(l) Is the threshold value of the jth node of the ith layer in the hidden layer.
By using the flow shown in fig. 4, after the initialization weight and the threshold value of the neural network optimization are obtained, the fine tuning training correction is continuously performed by using the BP algorithm, so that the prediction accuracy is improved. In the training process of the BP algorithm, output errors are propagated reversely, the errors are distributed to each layer of neural units, error signals of each layer of neural units are further obtained, the weight values of edges between the neural units in each layer are corrected, and finally an improved neural network is determined.
The fourth step: prediction output
Based on the third stepThe improved neural network is used for collecting the current operation data R of the power distribution and consumption systemx，
Figure GDA0002145252050000092
Inputting the current additional power instruction into an improved neural network, and further acquiring a current additional power instruction for intelligent operation control of the power distribution and consumption system based on data driving:
RY＝{ΔPdc,now,ΔPs1,now,ΔPs2,now}
although particular embodiments of the present invention have been described above, it will be appreciated by those skilled in the art that these are merely examples and that many variations or modifications may be made to these embodiments without departing from the principles and implementations of the invention, the scope of which is therefore defined by the appended claims.

Claims (4)

1. A data-driven intelligent operation control method for a power distribution and utilization system is characterized by comprising the following steps:
the first step is as follows: data acquisition and pretreatment, namely acquiring data from historical operating data of the multi-terminal AC/DC power distribution system, wherein the data are a plurality of groups of TxAnd TyA plurality of groups TxAnd TyAre combined into training sets T _ set, each set TxAnd TyForming a training data T, each training data T ═ Tx,TY}，TxIs the running data of a multi-terminal AC/DC distribution power system, TYIs corresponding to TxAn additional power command for the power distribution and utilization system; wherein:
TY＝{ΔPdc,ΔPs1,ΔPs2}
Figure FDA0002638251870000011
wherein i0 m,i0 s1,i0 s2,U0 dc,U0 s1,U0 s2,Udc,Us1,Us2Respectively representing master station linesThe steady state current value of the way, the steady state current value of the slave VSC1 line, the steady state current value of the slave VSC2 line, the steady state voltage value of the direct current bus, the steady state voltage value of the direct current side of the VSC1, the steady state voltage value of the direct current side of the VSC2, the value after the voltage change of the direct current bus, the value after the voltage change of the direct current side of the VSC1, the value after the voltage change of the direct current side of the VSC2, delta PdcFor additional power of energy storage devices, Δ Ps1Additional power, Δ P, for VSC1s2Additional power for VSC 2;
preprocessing the training data to obtain training neural network input values formed by corresponding training data;
the second step is that: unsupervised pre-training, training the training neural network input value correspondingly formed by each training data obtained in the first step by adopting Deep Belief Network (DBN), and completing i-learning by using a DBN deep learning structure and a self-adaptive weight learning rate0 m,i0 s1,i0 s2,U0 dc,U0 s1,U0 s2,Udc,Us1,Us2、ΔPdc、ΔPs1、ΔPs2Forming a neural network;
the third step: aiming at the neural network obtained in the second step, correcting the initialization weight and the threshold of the neural network by adopting a Particle Swarm Optimization (PSO) algorithm based on self-adaptive inertia weight factor updating and fitness value, and then carrying out supervised fine tuning by using a BP algorithm to obtain an improved neural network;
the fourth step: acquiring current operation data R of the power distribution and utilization system based on the improved neural network obtained in the third stepx，
Figure FDA0002638251870000012
Inputting the current additional power instruction into an improved neural network to obtain a current additional power instruction for intelligent operation control of the power distribution and consumption system based on data driving:
RY＝{ΔPdc,now,ΔPs1,now,ΔPs2,now}
wherein i0 m,now,i0 s1,now,i0 s2,now,U0 dc,now,U0 s1,now,U0 s2,now,Udc,now,Us1,now,Us2,nowThe steady state current value respectively represents the current state of the main station line, the steady state current value of the current state of the slave station VSC1 line, the steady state current value of the current state of the slave station VSC2 line, the steady state voltage value of the current state of the direct current bus, the direct current side steady state voltage value of the current state of the VSC1, the direct current side steady state voltage value of the current state of the VSC2, the current value of the direct current bus voltage, the current value of the direct current side voltage of the VSC1, the current value of the direct current side voltage of the VSC 36dc,nowAdditional power, Δ P, for the current state of the energy storage devices1,nowAdditional power, Δ P, for the current state of VSC1s2,nowIs the additional power for the current state of the VSC 2.
2. The intelligent operation control method based on the data-driven power distribution and utilization system as claimed in claim 1, wherein: in the first step, for a certain training data T, if
Figure FDA0002638251870000021
Wherein U isdc,min、Udc,max、Udc,ratedThe direct current bus voltage minimum value, the direct current bus voltage maximum value and the direct current bus voltage rated value are respectively obtained; u shapes1,min、Us1,max、Us1,ratedThe minimum value of the direct-current side voltage, the maximum value of the direct-current side voltage and the rated value of the direct-current side voltage of the VSC1 are respectively; u shapes2,min、Us2,max、Us2,ratedThe voltage minimum value, the voltage maximum value and the rated value of the direct current side of the VSC2 are respectively, and alpha is a starting threshold value;
then set in the training data T:
Figure FDA0002638251870000022
Figure FDA0002638251870000023
Figure FDA0002638251870000024
Figure FDA0002638251870000025
X5＝Udc
Figure FDA0002638251870000026
X7＝Us1
Figure FDA0002638251870000027
X9＝Us2
wherein, XiAnd (3) representing the ith input value of the training neural network, skipping the training data T if the ith input value does not meet the requirement, processing the 1 st to the last training data in the training set T _ set according to the steps, and acquiring the training neural network input values formed by corresponding training data.
3. The intelligent operation control method based on the data-driven power distribution and utilization system as claimed in claim 1, wherein: in the second step, the self-adaptive weight learning rate satisfies: when the updating direction of a certain weight value is consistent twice continuously, the step length is increased:
(m+1)＝(1-log(m))*(m)
when the updating direction of a certain weight value is inconsistent twice continuously, the step length is reduced:
(m+1)＝(1+log(m))*(m)
wherein (m +1) represents the weight learning rate at the m +1 th iteration, and (m) represents the weight learning rate at the m th iteration;
through the training process, the weights of the edges between different nodes in each hidden layer and the threshold values of each layer are obtained, and the corresponding neural network is obtained.
4. The intelligent operation control method based on the data-driven power distribution and utilization system as claimed in claim 1, wherein: the improved particle swarm PSO algorithm is realized based on a self-adaptive inertia weight factor updating method and a fitness value, wherein the formula of the self-adaptive inertia weight factor updating method is as follows:
Figure FDA0002638251870000031
wherein: ω (t +1) represents the PSO particle swarm inertial weight factor at the t-th iteration, ωmaxAnd ωminThe values of the upper boundary value and the lower boundary value of the inertia weight factor are 0.4 and 0.9; mmaxIs the maximum allowable iteration number; | v (t) | is the two-norm of the velocity vector at the t-th iteration; i V (t-1) I is the two-norm of the velocity vector at the time of the t-1 iteration;
the fitness value is composed of a prediction error absolute value sum of training data, a neural network connection weight square sum and a hidden layer threshold square sum, and the calculation formula is as follows:
Figure FDA0002638251870000032
in the formula, F represents the adaptability value of the PSO algorithm, N is the number of nodes of an output layer, and k1、k2Is a scale factor, ωij(l) As a weight of an edge between the jth node of the l layer and the ith node of the l-1 layer in the hidden layer, bj(l) Is the threshold value of the jth node of the ith layer in the hidden layer.
CN201910248085.8A 2019-03-29 2019-03-29 Intelligent operation control method of power distribution and utilization system based on data driving Active CN109995095B (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN201910248085.8A CN109995095B (en) 2019-03-29 2019-03-29 Intelligent operation control method of power distribution and utilization system based on data driving

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN201910248085.8A CN109995095B (en) 2019-03-29 2019-03-29 Intelligent operation control method of power distribution and utilization system based on data driving

Publications (2)

Publication Number Publication Date
CN109995095A CN109995095A (en) 2019-07-09
CN109995095B true CN109995095B (en) 2020-10-23

Family

ID=67131071

Family Applications (1)

Application Number Title Priority Date Filing Date
CN201910248085.8A Active CN109995095B (en) 2019-03-29 2019-03-29 Intelligent operation control method of power distribution and utilization system based on data driving

Country Status (1)

Country Link
CN (1) CN109995095B (en)

Families Citing this family (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN110401219B (en) * 2019-07-22 2020-10-09 中国科学院电工研究所 Intelligent operation method of distributed energy and user group alternating current-direct current power distribution and utilization system
CN111352344B (en) * 2020-02-28 2022-05-31 大连海事大学 Data-driven self-adaptive predictor of dynamic completely unknown system
CN112583039B (en) * 2020-12-21 2022-11-15 中国科学院电工研究所 Data-model combined driven AC/DC power distribution and utilization system coordination control method

Family Cites Families (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN101895125B (en) * 2010-08-06 2012-11-21 上海交通大学 Control method of light-type direct-current transmission system converter of offshore wind power station
CN106712103B (en) * 2017-03-14 2018-12-28 南京工程学院 A kind of micro-capacitance sensor voltage stabilizing control system and control method
CN107391852B (en) * 2017-07-26 2020-07-10 清华大学 Transient stability real-time evaluation method and device based on deep belief network
CN108767987A (en) * 2018-06-19 2018-11-06 清华大学 A kind of power distribution network and its micro-capacitance sensor protection and control system

Also Published As

Publication number Publication date
CN109995095A (en) 2019-07-09

Similar Documents

Publication Publication Date Title
Sun et al. A customized voltage control strategy for electric vehicles in distribution networks with reinforcement learning method
Zand et al. Using adaptive fuzzy logic for intelligent energy management in hybrid vehicles
CN109995095B (en) Intelligent operation control method of power distribution and utilization system based on data driving
Yang et al. Power management strategies for microgrid-A short review
CN108471143A (en) Micro-grid multi-energy method for optimizing scheduling based on positive and negative feedback particle cluster algorithm
Yin et al. Design of a novel smart generation controller based on deep Q learning for large-scale interconnected power system
CN110858718B (en) Alternating current micro-grid distributed event-driven frequency control method considering economy
Shang et al. Energy optimal dispatching of ship's integrated power system based on deep reinforcement learning
Zangeneh et al. A survey: Fuzzify parameters and membership function in electrical applications
Xiao et al. Ship energy scheduling with DQN-CE algorithm combining bi-directional LSTM and attention mechanism
Yin et al. Expandable deep width learning for voltage control of three-state energy model based smart grids containing flexible energy sources
Wei et al. Power balance control of RES integrated power system by deep reinforcement learning with optimized utilization rate of renewable energy
Maamir et al. Neural network power management for hybrid electric elevator application
CN110601246B (en) Direct-current micro-grid current sharing method based on radial basis function neural network prediction
CN117200213A (en) Power distribution system voltage control method based on self-organizing map neural network deep reinforcement learning
Ramasamy et al. Energy management in plugin hybrid electric vehicles with hybrid energy storage system using hybrid approach
CN116247742A (en) Power system source network load storage joint regulation and control method and device based on deep reinforcement learning
Wang et al. A real-time predictive energy management strategy of fuel cell/battery/ultra-capacitor hybrid energy storage system in electric vehicle
Irmak et al. Event‐triggered distributed secondary control for enhancing efficiency, reliability and communication in island mode DC microgrids
CN112510690B (en) Optimal scheduling method and system considering wind-fire-storage combination and demand response reward and punishment
CN112215462A (en) K-means clustering power grid adaptability evaluation method based on Mahalanobis distance
CN115528750B (en) Power grid safety and stability oriented data model hybrid drive unit combination method
CN111934307A (en) Flat operation control method and system for direct-current power distribution network
CN113629780B (en) Microgrid power converter control method, system, storage medium and device
CN115313510A (en) Adaptive reactive compensation photovoltaic inverter control method and system

Legal Events

Date Code Title Description
PB01 Publication
PB01 Publication
SE01 Entry into force of request for substantive examination
SE01 Entry into force of request for substantive examination
GR01 Patent grant
GR01 Patent grant