CN109784692B - Rapid safety constraint economic dispatching method based on deep learning - Google Patents

Rapid safety constraint economic dispatching method based on deep learning Download PDF

Info

Publication number
CN109784692B
CN109784692B CN201811631297.6A CN201811631297A CN109784692B CN 109784692 B CN109784692 B CN 109784692B CN 201811631297 A CN201811631297 A CN 201811631297A CN 109784692 B CN109784692 B CN 109784692B
Authority
CN
China
Prior art keywords
economic dispatching
safety
safety constraint
formula
deep learning
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
CN201811631297.6A
Other languages
Chinese (zh)
Other versions
CN109784692A (en
Inventor
杨知方
杨燕
余娟
代伟
雷星雨
向明旭
杨高峰
金黎明
古济铭
韩思维
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.)
Chongqing University
State Grid Corp of China SGCC
State Grid Chongqing Electric Power Co Ltd
Original Assignee
Chongqing University
State Grid Corp of China SGCC
State Grid Chongqing Electric Power Co Ltd
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 Chongqing University, State Grid Corp of China SGCC, State Grid Chongqing Electric Power Co Ltd filed Critical Chongqing University
Priority to CN201811631297.6A priority Critical patent/CN109784692B/en
Publication of CN109784692A publication Critical patent/CN109784692A/en
Application granted granted Critical
Publication of CN109784692B publication Critical patent/CN109784692B/en
Active legal-status Critical Current
Anticipated expiration legal-status Critical

Links

Images

Abstract

The invention discloses a fast safety constraint economic dispatching method based on deep learning, which mainly comprises the following steps: 1) and determining a deep neural network suitable for the safety constraint economic dispatching model. 2) The stack denoise autoencoder SDAE is trained. 3) And establishing a safety constraint economic dispatching model based on deep learning. 4) And (5) setting k to 1, inputting the operation condition of the power system into the deep neural network to obtain a working constraint set J of the safety constraint economic dispatching model(1). 5) Will constrain set J(1). And inputting the data into a safety constraint economic dispatching model to obtain a safety constraint economic dispatching scheme. 6) Carrying out N-1 inspection on the safety constraint economic dispatching scheme, and if a new constraint J exists(new)If k is k +1, the constraint set is updated to J(k)＝J(k)∪J(new)And returns to step 5. And if no new constraint exists, outputting a safety constraint economic scheduling scheme. The method can be widely applied to safety constraint economic dispatching analysis of various industries of the power system.

Description

Rapid safety constraint economic dispatching method based on deep learning
Technical Field
The invention relates to the field of electric power systems and automation thereof, in particular to a rapid safety constraint economic dispatching method based on deep learning.
Background
Safety constraint economic dispatch is an important tool for safe and economic operation of a power grid. But considering the line constraint of the N-1 fault can greatly increase the scale of the safety constraint economic dispatch model, which makes it a large-scale mathematical programming problem. This will bring an extensional problem to the solution of the safety constrained economic dispatch, which is difficult to solve even with commercial LP solvers. However, only a small number of all constraints are functional constraints and others are redundant constraints based on engineering experience.
Therefore, at present, the industry generally solves the unconstrained economic scheduling optimization problem, and iteratively adds constraints to the model for solving. In addition, in some foreign power structures, some most likely functional constraints are also added to the model in advance according to manual experience.
At present, a large number of scholars in the academic world propose scene screening methods to reduce redundant constraints and accelerate the solution of safety constraint economic dispatch, the methods mainly reduce the number of safety constraints by solving a plurality of small-scale optimization problems or establishing and solving another complex convex programming optimization problem, and the balance between the model solution precision and the redundancy constraint deletion number still has a space for improvement.
Disclosure of Invention
The present invention is directed to solving the problems of the prior art.
The technical scheme adopted for achieving the purpose of the invention is that the quick safety constraint economic dispatching method based on deep learning mainly comprises the following steps:
1) and determining a deep neural network suitable for the safety constraint economic dispatching model, namely a stack noise reduction automatic encoder SDAE.
The main steps for determining the deep neural network suitable for the safety-constrained economic dispatch are as follows:
1.1) build a stack de-noising autoencoder SDAE. The SDAE is formed by stacking n DAEs layer by layer.
Wherein the input layer of the first de-noising autoencoder DAE is denoted as Yl-1The middle layer is marked as YlOutput layer is marked as Zl。
Intermediate layer YlAs follows:
Figure BDA0001929056770000021
in the formula (I), the compound is shown in the specification,
Figure BDA0001929056770000022
representing the encoding function. R is an activation function. And theta is an encoding parameter. WlThe weight of the coding function of the l-th noise reduction auto-encoder DAE. blIs the offset of the coding function of the l-th noise reduction auto-encoder DAE.
Wherein the activation function R is as follows:
Figure BDA0001929056770000023
where x is the input to the neuron.
Output layer ZlAs follows:
Figure BDA0001929056770000024
in the formula (I), the compound is shown in the specification,
Figure BDA0001929056770000025
representing the decoding function. θ' is a decoding parameter. Wl' is the weight of the decoding function of the l-th noise reduction auto-encoder DAE. bl' is the offset of the decoding function of the l-th noise reduction auto-encoder DAE.
1.2) obtaining the operation condition of the power system, wherein the operation condition of the power system comprises load power PDAnd system topology.
1.3) the sample is preprocessed using equation 4.
Figure BDA0001929056770000026
Wherein v ismeanAnd vstdRespectively, the mean and standard deviation of vector V. V is the data that needs to be normalized, including the inputs and outputs of training samples, test samples, and test samples.
1.4) load power P after pretreatmentDInput into the SDAE to output power PG。
Output P of generatorGAs follows:
Figure BDA0001929056770000027
in the formula (I), the compound is shown in the specification,
Figure BDA0001929056770000028
the coding function of the nth noise reduction auto-encoder DAE.
2) And determining a deep learning strategy suitable for safety constraint economic dispatching, thereby training the stack noise reduction automatic encoder SDAE.
The main steps for training the SDAE are as follows:
2.1) carrying out unsupervised pre-training on the SDAE, and selecting a group of encoding parameters theta and decoding parameters theta' to ensure that the calculation parameter M reaches the minimum.
The parameter M is calculated as follows:
Figure BDA0001929056770000031
2.2) carrying out supervised fine tuning on the SDAE, namely selecting a coding parameter theta to ensure that a calculation parameter L reaches the minimum.
The parameter L is calculated as follows:
Figure BDA0001929056770000032
2.3) updating the coding parameter theta by utilizing an RMSprop learning algorithm, namely:
Figure BDA0001929056770000033
wherein the content of the first and second substances,
Figure BDA0001929056770000034
for the objective function O to the variable theta0At the t ththAnd (4) updating. As a Hadamard multiplier. ρ is a gradient accumulation index.
Figure BDA0001929056770000035
For the objective function O to the variableθAnd (4) updating.
Figure BDA0001929056770000036
Is a variable thetaoIteration at t-1. And r is the gradient. r istIs front tthThe gradient accumulated in the sub-iteration. r ist-1The gradient accumulated for the first t-1 iterations. Is a constant.
3) And establishing a safety constraint economic dispatching model based on deep learning based on the trained deep neural network.
The safety constraint economic dispatching model based on deep learning is as follows:
Figure BDA0001929056770000037
H1and H2Is a matrix of coefficients.
The constraints of the safety constraint economic dispatch model based on deep learning are shown in equations 10 to 12, respectively.
eGPG＝eDPD。 (10)
In the formula, eGAnd eDRepresenting vectors of all 1's.
Figure BDA0001929056770000038
In the formula (I), the compound is shown in the specification,
Figure BDA0001929056770000039
the branch power of branch ij is the branch power at the time of the c-th line fault, where c ═ 0 indicates a wireless line fault.
Figure BDA00019290567700000310
The lower branch power limit for branch ij at the time of the c-th line fault.
Figure BDA00019290567700000311
Is the branch power cap for branch ij at the time of the c-th line fault.
PG∈χG (12)
In the formula, xGIs the generator output set.
4) And (5) setting k to 1, inputting the operation condition of the power system into a deep neural network, and obtaining a working constraint set J of the safety constraint economic dispatching model based on deep learning(1)。
5) Will constrain set J(1). Input to depth-basedAnd obtaining a safety constraint economic dispatching scheme from the learned safety constraint economic dispatching model.
6) Carrying out N-1 inspection on the safety constraint economic dispatching scheme, and if a new constraint J exists(new)If k is k +1, the constraint set is updated to J(k)＝J(k)∪J(new)And returns to step 5. And if no new constraint exists, outputting a safety constraint economic scheduling scheme.
The technical effect of the present invention is undoubted. The present invention utilizes a stacked noise reduction auto-encoder (SDAE) to extract this nonlinear relationship based on historical operating data to "predict" a set of functional constraints on new system operating conditions. According to the invention, a deep learning technology is embedded into the existing scheduling framework, the calculation burden is transferred to off-line training, all the function constraint sets can be effectively mapped directly by the system operation conditions on line, and the safety constraint economic scheduling solving efficiency is improved.
The SDAE deep neural network-based functioning constraint identification method provided by the invention can quickly and accurately obtain a functioning constraint set of the safety constraint economic dispatching model from the system operation condition, and embodies the powerful approximation capability of the SDAE model to the nonlinear relation between the system operation condition and the generator output. The rapid safety constraint economic dispatching method based on the deep learning is embedded with the deep learning technology under the traditional optimization calculation framework, and the method can effectively improve the calculation speed under the condition of not influencing the calculation precision and the convergence. A new visual angle is provided for solving the problem of safety constraint economic dispatching.
The method can be widely applied to safety constraint economic dispatching analysis of various industries of the power system.
Drawings
FIG. 1 is a safety-constrained economic dispatch model based on deep learning;
fig. 2 is a diagram of a deep neural network DNN architecture.
Detailed Description
The present invention is further illustrated by the following examples, but it should not be construed that the scope of the above-described subject matter is limited to the following examples. Various substitutions and alterations can be made without departing from the technical idea of the invention and the scope of the invention is covered by the present invention according to the common technical knowledge and the conventional means in the field.
Example 1:
referring to fig. 1 and 2, a fast safety constraint economic dispatching method based on deep learning mainly includes the following steps:
1) a Deep Neural Network (DNN), i.e., a stacked noise reduction auto-encoder SDAE, is determined that is applicable to the safety-constrained economic dispatch model.
The main steps for determining the deep neural network suitable for the safety-constrained economic dispatch are as follows:
1.1) build Stack De-noising AutoCoder SDAE (stacked Denoising Autoencoders). The SDAE is formed by stacking n DAEs layer by layer.
Wherein the input layer of the first de-noising autoencoder DAE is denoted as Yl-1The middle layer is marked as YlOutput layer is marked as Zl。
Intermediate layer YlAs follows:
Figure BDA0001929056770000051
in the formula (I), the compound is shown in the specification,representing the encoding function. R is an activation function. θ is a coding parameter, and θ is { W, b }. WlThe weight of the coding function of the l-th noise reduction auto-encoder DAE. blIs the offset of the coding function of the l-th noise reduction auto-encoder DAE. W is the weight of the coding function. b is the bias of the coding function.
Wherein the activation function R is as follows:
Figure BDA0001929056770000053
in the formula, x represents the input of the neuron of the input layer, the intermediate layer or the output layer of the noise reduction auto-encoder DAE.
The piecewise Linear form of the ReLU (rectified Linear Unit) activation function is shown in a formula, and the ReLU activation function can effectively avoid the gradient disappearance phenomenon. Therefore, the ReLU function is chosen in the present invention as the activation function for SDAE.
Output Z of the output layerlAs follows:
Figure BDA0001929056770000054
in the formula (I), the compound is shown in the specification,
Figure BDA0001929056770000055
representing the decoding function. θ' is a decoding parameter. θ ' ═ W ', b ' }. Wl' is the weight of the decoding function of the l-th noise reduction auto-encoder DAE. bl' is the offset of the decoding function of the l-th noise reduction auto-encoder DAE. W' is the weight of the decoding function. b' is the offset of the decoding function.
1.2) obtaining the operating conditions of the power system, including the load power PDSystem topology, etc.
1.3) the sample is preprocessed using equation 4.
Figure BDA0001929056770000061
Wherein v ismeanAnd vstdRespectively, the mean and standard deviation of vector V. V is the data that needs to be normalized, including the inputs and outputs of training samples, test samples, and test samples. In this embodiment, the training samples are load power for training the SDAE, the testing samples are load power for testing the SDAE, and the testing samples are load power for testing the SDAE.
1.4) load power P after pretreatmentDInput into SDAE, and outputOutput P of starting motorG。
Output P of generatorGAs follows:
Figure BDA0001929056770000062
in the formula (I), the compound is shown in the specification,
Figure BDA0001929056770000063
the coding function of the nth noise reduction auto-encoder DAE.
Figure BDA0001929056770000064
The coding function of the 1 st noise reduction auto-encoder DAE.
2) And determining a deep learning strategy suitable for safety constraint economic dispatching, thereby training the stack noise reduction automatic encoder SDAE. The goal of deep neural network learning is to obtain the optimal coding parameters θ to extract the non-linear features between data.
The main steps for training the SDAE are as follows:
2.1) carrying out unsupervised pre-training on the SDAE, and selecting a group of encoding parameters theta and decoding parameters theta' to ensure that the calculation parameter M reaches the minimum.
The parameter M is calculated as follows:
Figure BDA0001929056770000065
2.2) carrying out supervised fine tuning on the SDAE, namely selecting a coding parameter theta to ensure that a calculation parameter L reaches the minimum.
The parameter L is calculated as follows:
Figure BDA0001929056770000066
2.3) updating the coding parameter theta by utilizing an RMSprop learning algorithm, namely:
Figure BDA0001929056770000071
wherein the content of the first and second substances,
Figure BDA0001929056770000072
for the objective function O to the variable theta0At the t ththAnd (4) updating. As a Hadamard multiplier. Eta is the learning rate. ρ is a gradient accumulation index.
Figure BDA0001929056770000073
For the objective function O to the variableθAnd (4) updating.
Figure BDA0001929056770000074
Is a variable thetaoIteration at t-1. And r is the gradient. r istIs front tthThe gradient accumulated in the sub-iteration. r ist-1The gradient accumulated for the first t-1 iterations. Is a constant. In the present embodiment, 10-8。
This embodiment adopts rmsprop (root mean square prediction) as the deep learning algorithm. The RMSprop learning algorithm learns training samples in batches. Each batch is trained in turn and parameters in the deep neural network are updated. In addition, the RMSprop learning algorithm makes the learning rate different for each parameter by using a moving average of the squared gradient. The algorithm can reduce the training stress and avoid local minima.
3) And establishing a safety constraint economic dispatching model based on deep learning based on the trained deep neural network.
The safety constraint economic dispatching model based on deep learning is as follows:
Figure BDA0001929056770000075
H1and H2Is a matrix of coefficients. The coefficient matrix is mainly composed of constraints J(k)And (6) determining.
The constraints of the safety constraint economic dispatch model based on deep learning are shown in equations 10 to 12, respectively.
eGPG＝eDPD。 (10)
In the formula, eGAnd eDRepresenting vectors of all 1, i.e., unit vectors.
Figure BDA0001929056770000076
In the formula (I), the compound is shown in the specification,
Figure BDA0001929056770000077
the branch power of branch ij is the branch power at the time of the c-th line fault, where c ═ 0 indicates a wireless line fault. J. the design is a square(k)A constraint set is represented.
Figure BDA0001929056770000078
The lower branch power limit for branch ij at the time of the c-th line fault.
Figure BDA0001929056770000079
Is the branch power cap for branch ij at the time of the c-th line fault.
PG∈χG (12)
In the formula, xGIs the generator output set.
4) And (5) setting k to 1, inputting the operation condition of the power system into a deep neural network, and obtaining a working constraint set J of the safety constraint economic dispatching model based on deep learning(1)。
The operating conditions include load power and topology, etc. The present embodiment represents the operating conditions only in terms of load power.
5) Will constrain set J(1). Inputting the data into a safety constraint economic dispatching model based on deep learning to obtain a safety constraint economic dispatching scheme, namely meeting a constraint set J(k)Is/are as follows
Figure BDA0001929056770000081
6) Carrying out N-1 inspection on the safety constraint economic dispatching scheme, and if a new constraint J exists(new)If k is k +1, the constraint set is updated to J(k)＝J(k)∪J(new)And returns to step 5. And if no new constraint exists, outputting a safety constraint economic scheduling scheme.
The N-1 test principle is a criterion for determining the safety of the power system, and is also called a single failure safety criterion. According to the principle, after any independent element (generator, transmission line, transformer and the like) in N elements of the power system is cut off due to a fault, no power failure of a user due to overload tripping of other lines is caused, the stability of the system is not damaged, and accidents such as voltage breakdown and the like do not occur.
Example 2:
an experiment for verifying a fast safety constraint economic dispatching method based on deep learning mainly comprises the following steps:
1) training sample acquisition and preprocessing
In this embodiment, an IEEE-118-Washington system is adopted for simulation. The load active power PD is sampled in a normal distribution to represent different operating conditions, wherein the load mean is the value in the IEEE118 standard system and the standard deviation is 10% of the mean. Solving the safety constraint economic dispatching model corresponding to each sampling state by a large-scale optimizer Gurobi solver to obtain the optimal generator output PG. Then, the active power of the load nodes in all sampling states is used as a training sample input X, and the output of the generator is used as a training sample output Y. Finally, the training samples are preprocessed by equation (4).
(2) Deep neural network initialization suitable for safety-constrained economic dispatch
Dividing the training samples into 100 batches according to the capacity of the training samples; according to the scale and the complexity of the electric power system to be solved, the number l of layers of the SDAE optimal power flow model is set to be 6, the number of neurons in each layer is respectively 118, 200 and 19, and other hyper-parameters are shown in Table 1.
TABLE 1 deep neural network hyper-parameters
Figure BDA0001929056770000091
(3) Deep neural network unsupervised pre-training suitable for safety-constrained economic dispatch
First, P is input using training samplesDAn objective function M (Z) of the first-level DAE training is constructed according to equation (6)1,PD) (ii) a Then, according to the formula (2) and the formula (8), the RMSProp learning algorithm is used for iteratively solving the optimal parameter W of the first layer DAE1、b1、W′1、b′1. Then, the intermediate layer output of the first layer DAE is obtained from the formula (1) and the formula (3) and is used as the input of the second layer DAE, and the loss function M (Z) of the second layer DAE is constructed in the same way2,Z1) And updating parameters by the same method, and by analogy, solving the optimal coding parameter theta of each layer of DAE layer by layer from bottom to top, namely { W, b }, and taking the optimal coding parameter as an initial value of the next-stage supervised fine tuning.
(4) Deep neural network supervised fine tuning suitable for safety constraint economic dispatch
First, P is input using training samplesDAnd an output PGAn objective function L (P) of the SDAE supervised training process is constructed according to equation (6)G,PD) (ii) a Then, according to the formulas (2) and (8), iteratively solving all optimal coding parameters theta of the SDAE by using a RMSProp learning algorithm, wherein the optimal coding parameters theta are { W, b }; therefore, the deep neural network training suitable for the safety constraint economic dispatching is completed.
(5) Safety constrained economic dispatch fast calculation
Inputting the test samples into the SDAE optimal model trained in the step (4), wherein the model can be directly mapped by formula (5) to obtain the optimal power generation output of all the test samples considering the safety of N-1; performing N-1 inspection according to the output of the generator to determine an action constraint set; adding the action constraint set into a safety constraint economic dispatching model, and performing optimization calculation as shown in formulas (9) - (12); and substituting the optimization result into the N-1 test again to check whether new constraints exist, if not, finishing the calculation, otherwise, adding the new constraints into the safety constraint economic dispatching model for optimization again.
The specific simulation results are as follows:
I) safety constraint economic dispatching calculation comparison method
The safety constraint economic dispatch comparison method in simulation comprises the following steps of M0-M1:
m0 Industrial Process.
M1 fast safety constraint economic dispatching method based on deep learning.
The simulation results of this example were tested in the hardware environment of Intel (R) core (TM) i5-7200U CPU @2.50GHz 2.71GHz, 16GB RAM.
II) analysis of the recognition accuracy of the constraints of the deep learning functioning
In this embodiment, in order to verify the overall accuracy of the deep neural network identification functioning constraint, 2000 test samples are randomly extracted by an example, and through testing, 64,356,000 constraints are totally contained in the 2000 samples, wherein the functioning constraint set is 112,498, and the accuracy of determining whether the constraint set is functioning in the method provided by the present invention reaches 96.6%. Therefore, the SDAE effectively extracts the nonlinear characteristics between the system operation condition and the optimal generated output by virtue of the deep stack structure and the encoding and decoding process of the SDAE, and realizes high-precision and rapid mapping from the system operation condition to the action constraint set. Therefore, the deep neural network function constraint pre-identification method provided by the invention has higher precision.
III) probabilistic optimal power flow on-line algorithm calculation performance analysis
Table 2 lists the number of cycles for the M0 and M1 methods to compute safety constrained economic dispatch, 95.3% of which converged at once in 2000 test samples due to the high accuracy of the deep neural network's identification of the acting constraints. Table 3 lists the calculated time from the M1 method in a scenario where M0 solves the safety constrained economic dispatch problem for four loop iterations. As can be seen from table 3, the calculation of the safety-constrained economic scheduling problem by the M0 method takes 13.2 seconds, whereas the M1 method requires only 4.0 seconds, and the calculation speed is 3.3 times that of the M0 method. Therefore, the method for computing the safety constraint economic dispatching problem can effectively improve the computing speed, and the computing precision and the convergence are not influenced. In addition, it is noted that the single moment is considered in the embodiment of the present invention, and if a large-scale practical system and a 24-moment combined scheduling scenario are considered, the advantage of the fast safety constraint economic scheduling method based on deep learning will be more obvious.
TABLE 2M 0-M1 calculate number of loop iterations for safety-constrained economic scheduling
Figure BDA0001929056770000101
Table 3M0-M1 calculate time comparisons for safety-constrained economic schedules
Method of producing a composite material M0 M1
Time (seconds) 13.2 4.0
From the experimental results, it can be seen that: the rapid safety constraint economic dispatching method based on deep learning provided by the invention can approach the complex nonlinear relation between the system operation condition and the optimal output of the generator with very high precision, thereby realizing the efficient identification of the system operation condition with function constraint and having the characteristic of effectively improving the precision without loss and speed. In addition, the applicability of the invention will be more significant for large-scale systems and multi-period scheduling problems.
In conclusion, the invention provides a fast safety constraint economic dispatching method based on deep learning, which can transfer the calculation pressure of the safety constraint economic dispatching problem to linear training by depending on the existing calculation resources and historical operation data. By embedding the deep neural network into the existing computation framework of the safety constraint economic scheduling problem, the computation precision and the convergence can be not lost, the iteration times can be effectively reduced, and the computation efficiency can be improved. The effectiveness of the method provided by the invention is verified through example simulation analysis. Therefore, the method can provide technical support for efficient optimization of power system safety constraint economic dispatch.

Claims (3)

1. A fast safety constraint economic dispatching method based on deep learning is characterized by comprising the following steps:
1) determining a deep neural network suitable for a safety constraint economic dispatching model, namely a stack noise reduction automatic encoder SDAE;
2) determining a deep learning strategy suitable for safety constraint economic dispatching, and training a stack noise reduction automatic encoder SDAE;
the steps for training the stack denoise autoencoder SDAE are as follows:
2.1) carrying out unsupervised pre-training on the SDAE, and selecting a group of encoding parameters theta and decoding parameters theta' to ensure that the calculation parameter M reaches the minimum;
the parameter M is calculated as follows:
Figure FDA0002707867100000011
in the formula (I), the compound is shown in the specification,
Figure FDA0002707867100000012
representing an encoding function; y isl-1Is the input layer of the l-th de-noising auto-encoder DAE;
Figure FDA0002707867100000013
representing a decoding function;
2.2) carrying out supervised fine adjustment on the SDAE, namely selecting a coding parameter theta to ensure that a calculation parameter L reaches the minimum;
the parameter L is calculated as follows:
Figure FDA0002707867100000014
in the formula, PGOutputting power for the generator;
Figure FDA0002707867100000015
an encoding function for the nth noise reduction auto encoder DAE; pDThe load power after the pretreatment;
2.3) updating the coding parameter theta by utilizing an RMSprop learning algorithm, namely:
Figure FDA0002707867100000016
wherein the content of the first and second substances,
Figure FDA0002707867100000017
for the objective function O to the variable thetaoUpdating in the t-th iteration; as a Hadamard multiplier; eta is the learning rate; ρ is a gradient accumulation index;
Figure FDA0002707867100000018
updating a variable theta for the objective function O;
Figure FDA0002707867100000019
is a variable thetaoIteration at t-1; r is the gradient; r istGradients accumulated for the previous t iterations; r ist-1Gradients accumulated for the first t-1 iterations; is a constant;
3) establishing a safety constraint economic dispatching model based on deep learning based on the trained deep neural network;
4) inputting the operating condition of the power system into the deep neural network by setting the iteration number k to be 1 to obtain the basisAction constraint set J of deep learning safety constraint economic dispatching model(1)；
5) Will constrain set J(1)Inputting the data into a safety constraint economic dispatching model based on deep learning to obtain a safety constraint economic dispatching scheme;
6) carrying out N-1 inspection on the safety constraint economic dispatching scheme, and if a new constraint J exists(new)If k is k +1, the constraint set is updated to J(k)＝J(k)∪J(new)And returning to the step 5); and if no new constraint exists, outputting a safety constraint economic scheduling scheme.
2. The fast safety constraint economic dispatching method based on deep learning as claimed in claim 1, characterized in that: the steps of determining the deep neural network suitable for the safety-constrained economic dispatch are as follows:
1) establishing a stack noise reduction automatic encoder SDAE; the SDAE is formed by stacking n DAEs layer by layer;
wherein the input layer of the first de-noising autoencoder DAE is denoted as Yl-1The middle layer is marked as YlOutput layer is marked as Zl；
Intermediate layer YlAs follows:
Figure FDA0002707867100000021
in the formula (I), the compound is shown in the specification,
Figure FDA0002707867100000022
representing an encoding function; r is an activation function; theta is a coding parameter; wlWeights for the coding function of the l-th de-noising autoencoder DAE; blAn offset of the coding function for the l-th denoised auto-encoder DAE;
wherein the activation function R is as follows:
Figure FDA0002707867100000023
wherein x is the input to the neuron;
output layer ZlAs follows:
Figure FDA0002707867100000024
in the formula (I), the compound is shown in the specification,
Figure FDA0002707867100000025
representing a decoding function; theta' is a decoding parameter; wl' weight of decoding function of the l-th noise reduction auto-encoder DAE; b'lAn offset of the decoding function for the l-th denoised auto-encoder DAE;
2) acquiring the operating conditions of the power system; the power system operating condition comprises load power PDAnd power system topology;
3) preprocessing a sample by using a formula (7);
Figure FDA0002707867100000031
wherein v ismeanAnd vstdRespectively are the mean value and standard deviation of the vector V, and V is data needing normalization processing, including input and output of training samples, test samples and test samples;
4) load power P after pretreatmentDInput into the stack noise reduction automatic encoder SDAE to output the output P of the generatorG；
Output P of generatorGAs follows:
Figure FDA0002707867100000032
in the formula (I), the compound is shown in the specification,
Figure FDA0002707867100000033
the coding function of the nth noise reduction auto-encoder DAE.
3. The fast safety-constrained economic dispatching method based on deep learning as claimed in claim 1 or 2, characterized in that the safety-constrained economic dispatching model based on deep learning is as follows:
Figure FDA0002707867100000034
in the formula, H1And H2Is a coefficient matrix; pGOutputting power for the generator;
the constraints of the safety constraint economic dispatching model based on deep learning are respectively shown in the formula (10) to the formula (12);
eGPG＝eDPD； (10)
in the formula, eGAnd eDRepresents a unit vector;
Figure FDA0002707867100000035
in the formula (I), the compound is shown in the specification,
Figure FDA0002707867100000036
branch power for branch ij at the time of the c-th line fault, where c ═ 0 indicates a wireless line fault;
Figure FDA0002707867100000037
the lower limit of the branch power of the branch ij when the c line fails;
Figure FDA0002707867100000038
is the branch power upper limit of branch ij at the time of the c-th line fault;
PG∈χG (12)
in the formula, xGIs the generator output set.
CN201811631297.6A 2018-12-29 2018-12-29 Rapid safety constraint economic dispatching method based on deep learning Active CN109784692B (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN201811631297.6A CN109784692B (en) 2018-12-29 2018-12-29 Rapid safety constraint economic dispatching method based on deep learning

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN201811631297.6A CN109784692B (en) 2018-12-29 2018-12-29 Rapid safety constraint economic dispatching method based on deep learning

Publications (2)

Publication Number Publication Date
CN109784692A CN109784692A (en) 2019-05-21
CN109784692B true CN109784692B (en) 2020-11-24

Family

ID=66497987

Family Applications (1)

Application Number Title Priority Date Filing Date
CN201811631297.6A Active CN109784692B (en) 2018-12-29 2018-12-29 Rapid safety constraint economic dispatching method based on deep learning

Country Status (1)

Country Link
CN (1) CN109784692B (en)

Families Citing this family (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN110556823B (en) * 2019-08-15 2023-11-14 中国南方电网有限责任公司 Model dimension reduction based safety constraint unit combination rapid calculation method and system
CN110929989B (en) * 2019-10-29 2023-04-18 重庆大学 N-1 safety checking method with uncertainty based on deep learning
CN110991731A (en) * 2019-11-28 2020-04-10 中国南方电网有限责任公司 Electric power real-time market constraint self-identification clearing method and system based on deep learning
CN110991741B (en) * 2019-12-02 2022-07-12 中国南方电网有限责任公司 Section constraint probability early warning method and system based on deep learning
CN111612213B (en) * 2020-04-10 2023-10-10 中国南方电网有限责任公司 Section constraint intelligent early warning method and system based on deep learning
CN113761788A (en) * 2021-07-19 2021-12-07 清华大学 SCOPF rapid calculation method and device based on deep learning

Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN107491867A (en) * 2017-08-07 2017-12-19 国电南瑞科技股份有限公司 It is a kind of for the multicycle send out defeated change repair schedule Security Checking and appraisal procedure
CN108054757A (en) * 2017-12-22 2018-05-18 清华大学 A kind of embedded idle and voltage N-1 Close loop security check methods
CN108304623A (en) * 2018-01-15 2018-07-20 重庆大学 A kind of Probabilistic Load Flow on-line calculation method based on storehouse noise reduction autocoder
WO2018180386A1 (en) * 2017-03-30 2018-10-04 国立研究開発法人産業技術総合研究所 Ultrasound imaging diagnosis assistance method and system

Family Cites Families (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20130060397A1 (en) * 2011-09-07 2013-03-07 General Electric Company Management of power distribution constraints
US9705336B2 (en) * 2013-11-14 2017-07-11 Abb Research Ltd. Method and apparatus for security constrained economic dispatch in hybrid power systems

Patent Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
WO2018180386A1 (en) * 2017-03-30 2018-10-04 国立研究開発法人産業技術総合研究所 Ultrasound imaging diagnosis assistance method and system
CN107491867A (en) * 2017-08-07 2017-12-19 国电南瑞科技股份有限公司 It is a kind of for the multicycle send out defeated change repair schedule Security Checking and appraisal procedure
CN108054757A (en) * 2017-12-22 2018-05-18 清华大学 A kind of embedded idle and voltage N-1 Close loop security check methods
CN108304623A (en) * 2018-01-15 2018-07-20 重庆大学 A kind of Probabilistic Load Flow on-line calculation method based on storehouse noise reduction autocoder

Non-Patent Citations (2)

* Cited by examiner, † Cited by third party
Title
含电转气的电_气互联系统可靠性评估;余娟;《中国电机工程学报》;20180205;第38卷(第03期);全文 *
考虑含有新能源的电力系统安全约束经济调度;白立鹏;《中国优秀硕士学位论文全文数据库 工程科技Ⅱ辑》;20180315(第03期);全文 *

Also Published As

Publication number Publication date
CN109784692A (en) 2019-05-21

Similar Documents

Publication Publication Date Title
CN109784692B (en) Rapid safety constraint economic dispatching method based on deep learning
Wu et al. Data-driven remaining useful life prediction via multiple sensor signals and deep long short-term memory neural network
WO2020244134A1 (en) Multi-task feature sharing neural network-based intelligent fault diagnosis method
CN108445752B (en) Random weight neural network integrated modeling method for self-adaptively selecting depth features
WO2021257128A2 (en) Quantum computing based deep learning for detection, diagnosis and other applications
Dehghani et al. Adaptive network reliability analysis: Methodology and applications to power grid
CN111242351A (en) Tropical cyclone track prediction method based on self-encoder and GRU neural network
CN110851654A (en) Industrial equipment fault detection and classification method based on tensor data dimension reduction
Mylonas et al. Conditional variational autoencoders for probabilistic wind turbine blade fatigue estimation using Supervisory, Control, and Data Acquisition data
CN113505458A (en) Cascading failure key trigger branch prediction method, system, equipment and storage medium
Chen et al. Research on wind power prediction method based on convolutional neural network and genetic algorithm
Huang et al. An adversarial learning approach for machine prognostic health management
CN104732067A (en) Industrial process modeling forecasting method oriented at flow object
CN113947182A (en) Traffic flow prediction model construction method based on double-stage stack graph convolution network
CN112989976A (en) Equipment failure mode prediction method based on double-depth learning model
CN111784061A (en) Training method, device and equipment for power grid engineering cost prediction model
CN116929815A (en) Equipment working state monitoring system and method based on Internet of things
Su et al. Generative adversarial networks for gearbox of wind turbine with unbalanced data sets in fault diagnosis
CN115481788B (en) Phase change energy storage system load prediction method and system
Qiu et al. Air traffic flow of genetic algorithm to optimize wavelet neural network prediction
Gao et al. RUL prediction for IMA based on deep regression method
Ge et al. Remaining useful life prediction using deep multi-scale convolution neural networks
CN114298413A (en) Hydroelectric generating set runout trend prediction method
WO2014173271A1 (en) Optimization method and system for the number of monitoring units of digital man-machine interface
Ding et al. Evolving neural network using hybrid genetic algorithm and simulated annealing for rainfall-runoff forecasting

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