CN112508239A - Energy storage output prediction method based on VAE-CGAN - Google Patents
Energy storage output prediction method based on VAE-CGAN Download PDFInfo
- Publication number
- CN112508239A CN112508239A CN202011315545.3A CN202011315545A CN112508239A CN 112508239 A CN112508239 A CN 112508239A CN 202011315545 A CN202011315545 A CN 202011315545A CN 112508239 A CN112508239 A CN 112508239A
- Authority
- CN
- China
- Prior art keywords
- data
- sequence
- discriminator
- generator
- sample data
- 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.)
- Pending
Links
- 238000004146 energy storage Methods 0.000 title claims abstract description 51
- 238000000034 method Methods 0.000 title claims abstract description 31
- 238000012549 training Methods 0.000 claims abstract description 17
- 238000013528 artificial neural network Methods 0.000 claims description 54
- 210000002569 neuron Anatomy 0.000 claims description 50
- 238000004088 simulation Methods 0.000 claims description 31
- 238000009826 distribution Methods 0.000 claims description 28
- 238000010248 power generation Methods 0.000 claims description 21
- 238000013507 mapping Methods 0.000 claims description 11
- 239000011159 matrix material Substances 0.000 claims description 11
- 238000012937 correction Methods 0.000 claims description 6
- 238000005457 optimization Methods 0.000 claims description 5
- 238000012217 deletion Methods 0.000 claims description 4
- 230000037430 deletion Effects 0.000 claims description 4
- 238000005315 distribution function Methods 0.000 claims description 4
- 238000012545 processing Methods 0.000 claims description 4
- 238000004364 calculation method Methods 0.000 claims description 3
- 108700041286 delta Proteins 0.000 claims description 3
- 238000005070 sampling Methods 0.000 claims description 3
- 230000008569 process Effects 0.000 abstract description 7
- 230000002349 favourable effect Effects 0.000 abstract 1
- 230000006870 function Effects 0.000 description 9
- 238000004422 calculation algorithm Methods 0.000 description 5
- 230000001965 increasing effect Effects 0.000 description 4
- 230000000694 effects Effects 0.000 description 2
- 230000002159 abnormal effect Effects 0.000 description 1
- 230000003044 adaptive effect Effects 0.000 description 1
- 230000009286 beneficial effect Effects 0.000 description 1
- 230000008859 change Effects 0.000 description 1
- 238000011161 development Methods 0.000 description 1
- 238000010586 diagram Methods 0.000 description 1
- 230000002708 enhancing effect Effects 0.000 description 1
- 238000012986 modification Methods 0.000 description 1
- 230000004048 modification Effects 0.000 description 1
- 238000010606 normalization Methods 0.000 description 1
- 230000000737 periodic effect Effects 0.000 description 1
- 230000006403 short-term memory Effects 0.000 description 1
- 239000002699 waste material Substances 0.000 description 1
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06Q—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES; SYSTEMS OR METHODS SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES, NOT OTHERWISE PROVIDED FOR
- G06Q10/00—Administration; Management
- G06Q10/04—Forecasting or optimisation specially adapted for administrative or management purposes, e.g. linear programming or "cutting stock problem"
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/044—Recurrent networks, e.g. Hopfield networks
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/045—Combinations of networks
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06Q—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES; SYSTEMS OR METHODS SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES, NOT OTHERWISE PROVIDED FOR
- G06Q10/00—Administration; Management
- G06Q10/06—Resources, workflows, human or project management; Enterprise or organisation planning; Enterprise or organisation modelling
- G06Q10/067—Enterprise or organisation modelling
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06Q—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES; SYSTEMS OR METHODS SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES, NOT OTHERWISE PROVIDED FOR
- G06Q50/00—Information and communication technology [ICT] specially adapted for implementation of business processes of specific business sectors, e.g. utilities or tourism
- G06Q50/06—Energy or water supply
Landscapes
- Engineering & Computer Science (AREA)
- Business, Economics & Management (AREA)
- Physics & Mathematics (AREA)
- Theoretical Computer Science (AREA)
- Human Resources & Organizations (AREA)
- Strategic Management (AREA)
- Economics (AREA)
- General Physics & Mathematics (AREA)
- Health & Medical Sciences (AREA)
- Marketing (AREA)
- Entrepreneurship & Innovation (AREA)
- Tourism & Hospitality (AREA)
- General Health & Medical Sciences (AREA)
- General Business, Economics & Management (AREA)
- Data Mining & Analysis (AREA)
- Biomedical Technology (AREA)
- Software Systems (AREA)
- General Engineering & Computer Science (AREA)
- Computing Systems (AREA)
- Development Economics (AREA)
- Molecular Biology (AREA)
- Game Theory and Decision Science (AREA)
- Evolutionary Computation (AREA)
- Computational Linguistics (AREA)
- Biophysics (AREA)
- Mathematical Physics (AREA)
- Operations Research (AREA)
- Quality & Reliability (AREA)
- Artificial Intelligence (AREA)
- Life Sciences & Earth Sciences (AREA)
- Public Health (AREA)
- Water Supply & Treatment (AREA)
- Primary Health Care (AREA)
- Educational Administration (AREA)
- Management, Administration, Business Operations System, And Electronic Commerce (AREA)
Abstract
A VAE-CGAN-based energy storage output prediction method comprises the following steps: collecting historical output data of an energy storage power station; obtaining characteristic information of historical output data through training of a VAE model; inputting the characteristic information into a generator and a discriminator, setting noise information and inputting historical output data of the energy storage power station into the discriminator; and the generator generates sample data and inputs the sample data into the discriminator, the discriminator judges the source of the generated sample data according to the condition information and the real sample data, and the discriminator can modify parameters and repeat the generation and discrimination processes until the discriminator can not sample the data source, so that the stronger generator can predict the output condition of the energy storage device. Compared with the prior art, the model constructs a continuous condition input space, and the method can generate data according to condition information in a historical scene and accurately generate according to unknown condition information, and is favorable for accurate output modeling of the stored energy in an uncertain scene.
Description
The technical field is as follows:
the invention belongs to the technical field of power distribution network energy storage output modeling, and particularly relates to a VAE-CGAN-based energy storage output prediction method.
Background art:
in recent years, new energy power generation is continuously developed, and the development of forms of solar power generation, wind power, tidal power and the like also brings new challenges to a power grid. For example, the forms of solar power generation, tidal power generation and the like are influenced by conditions such as weather and tidal cycle, and the power generation is intermittent; and wind power generation is also influenced by different wind power in different seasons, so that the power generation has uncertainty. As the proportion of new energy power generation is increasing, the influence of instability of new energy power generation on the power system is increasing. Naturally also the modeling of the output of the stored energy is affected.
The energy storage device has the effect of 'peak clipping and valley filling', can cope with the load in the peak period of the load, and reduces the peak-valley difference. The power grid enterprise can obtain more peak load benefits while adjusting the peak and relieving the power supply pressure. The output of the energy storage device is influenced by factors such as new energy power generation, weather, seasons and the like; meanwhile, although the electrical load is basically periodic, the electrical load still fluctuates continuously, and therefore, the determination of the energy storage output is influenced. The significance of the energy storage output modeling is that the electric energy can be dispatched in a more reasonable operation mode, the cost of a power grid enterprise can be saved, the profit can be improved, and the utilization rate of new energy can be improved.
For the reasons mentioned above, the determination of the energy storage capacity needs to be faced with the influence of uncertain external conditions. Due to the influence of unstable factors in the aspects of new energy power generation and load, the method establishes a continuous condition input space by using a VAE (variable auto-encoder) model, and inputs a CGAN (Conditional access adaptive network) model to establish an energy storage output model. Therefore, the energy storage output uncertainty modeling based on the VAE-CGAN can not only generate condition information in a historical scene, but also train characteristic information of the energy storage output uncertainty modeling by using a VAE model in the face of abnormal external conditions and perform modeling, and can accurately and comprehensively model the output of the energy storage while ensuring the accuracy.
The invention content is as follows:
the invention aims to determine the energy storage output force under different loads and generated energy under the condition of unstable generated energy of new energy, thereby enhancing the peak clipping and valley filling effects, reducing the waste conditions such as air abandon and the like and improving the economy.
The invention specifically adopts the following technical scheme:
a VAE-CGAN-based energy storage output prediction method is characterized by comprising the following steps:
step 1: collecting historical operation data of the energy storage device, wherein the historical operation data comprises battery voltage, battery ampere hours, discharge multiplying power, battery operation voltage, traditional energy power generation ratio and new energy power generation ratio;
step 2: directly inputting the historical operating data acquired in the step 1 into a VAE model for training to generate data characteristic information, namely generating new data containing information of the historical operating data;
and step 3: constructing a generator and a discriminator based on the CGAN, wherein the generator is used for generating simulation sample data, and the discriminator is used for discriminating the sources of the simulation sample data and real sample data;
and 4, step 4: inputting the characteristic information obtained in the step 2 as condition information into the generator and the discriminator, and performing countermeasure training on the generator and the discriminator, namely, the discriminator receives the simulation sample data and the real sample data acquired in the step 1 and judges the sources of the two data;
and 5: updating the parameters of the generator and the discriminator, namely optimizing the generator and the discriminator;
step 6: adding 1 to the iteration times, returning to the step 5 until the discriminator cannot distinguish the sample data source, and outputting a generator;
and 7: and (4) acquiring real-time operation data of the energy storage device, inputting the real-time operation data into the generator output in the step (6), and predicting the energy storage output condition of the energy storage device.
The invention further adopts the following preferred technical scheme:
the step 2 comprises the following steps:
step 201: constructing an encoder by adopting a layer of LSTM neural network; a decoder is constructed by adopting a layer of LSTM neural network and a layer of full connection layer;
step 202: the encoder receives a first high-dimensional data sequence X ═ X of historical operating data1,x2,…xn]Then, mapping a high-dimensional data sequence X of the historical operating data into a mean vector mu with the length of m and a standard deviation vector sigma with the length of m;
step 203: the encoder calculates a mean vector mu, a standard deviation vector sigma and a parameter sequence delta [ delta 1, delta 2, … delta m ═ according to the mean vector mu and the standard deviation vector sigma]And calculating a hidden variable sequence Z ═ Z by the following formula1,z2,…zm];
zi=μi+δi·exp(σi)
Wherein z isiFor the ith value in the hidden variable sequence Z,μiIs the ith value, σ, in the mean vector μiIs the ith value in the standard deviation vector σ, δ i is a parameter obtained by random sampling in a sample set subject to bernoulli distribution, and subject to a standard normal distribution, i.e., δ i to N (0,1), i ═ 1,2, …, m; step 204: inputting the hidden variable sequence Z obtained by calculation in the step 203 into a decoder, and restoring the hidden variable sequence Z into a second high-dimensional data sequence X ', wherein the second high-dimensional data sequence X' is a characteristic information sequence.
In step 2, the following objective function is adopted for training:
wherein L is1Is a self-encoding reconstruction error, L2Is KL divergence, P (x | Z) is a prior distribution of hidden variables Z representing a decoder in the VAE, q (Z | x) is a posterior distribution of values Z in a sequence of hidden variables derived from any value x in the real data sequence representing an encoder in the VAE;
and when the self-coding reconstruction error value is maximum and the KL divergence value is minimum, finishing training to obtain the characteristic information.
In the step 2, the battery voltage, the battery ampere hour and the discharge multiplying power are trained by adopting the same VAE model;
and the traditional energy power generation ratio and the new energy power generation ratio are trained by adopting independent VAE models respectively.
In step 4, random noise information Z is setnoiseAnd input into a generator together with the condition information, the generator utilizes the nonlinear mapping capability of the neural network to give the noise information znoiseAnd condition information Y into simulation sample data G (z)noise|y)。
In step 4, the generator generating the simulation sample comprises the following steps:
step 401: input random noise sequence ZnoiseWhen the length is n and the length of the conditional information sequence Y is m, the generator establishes a full connection layer to enable the generator to be connected with the random noise sequence ZnoiseThe number of the neurons of the full connection layer corresponding to the condition information sequence Y and the random noise sequence Z are respectivelynoiseThe length of the conditional information sequence Y is consistent with that of the conditional information sequence Y;
step 402: the data of the fully connected layer is corrected such that its mean value is approximately 0 and its variance is approximately 1, i.e., towards a standard normal distribution N (0, 1).
Step 403: random noise Z to be correctednoiseSplicing the correction sequence with the condition information Y correction sequence to form a spliced sequence with the length of (n + m)/10;
step 404: multiplying the same appearance probability p by (n + m)/10 neurons in the spliced sequence with the length of (n + m)/10 obtained in the previous layer to be 0.5;
step 405: in step 404, half of the neurons are temporarily deleted in the current generation, and the undeleted neurons are output from the generator as the simulation sample data.
In step 4, the discriminator discriminating the source of the input data comprises the following steps:
step 406: the discriminator receives the condition information Y, the real sample data sequence X and the simulation sample data sequence generated by the generator;
step 407: the discriminator respectively generates a first hidden layer for the condition information sequence Y, the real sample data sequence X and the simulation sample data sequence, wherein the number of the neurons of the hidden layer is i, and a first weight matrix W is establishedmaxout1Calculating the numerical value of the neuron in the first hidden layer by the following formula;
t’i=wi1×t1+wi2×t2+....+win×tn
wherein, wilIs a weight matrix Wmaxout1The 1 st element of the ith row, t'iIs the value of the ith neuron in the first hidden layer, tiIs the ith data of the input sequence.
Step 408: dividing neurons in a first hidden layer of a condition information sequence Y, a first hidden layer of a real sample data sequence X and a first hidden layer of a simulation sample data sequence into s groups, selecting the neuron with the largest value from each group of the groups as the output of the first hidden layer, and generating a condition information neural network layer, a real sample data neural network layer and a simulation sample data neural network layer, wherein the neural network layer contains s neurons;
step 409: splicing the generation condition information neural network layer, the real sample data neural network layer and the simulation sample data neural network layer to obtain a first neural network layer, wherein the first neural network layer is provided with 3s neurons;
step 410: setting the neuron deletion probability to be 0.5, and processing the first neural network layer in the step 409 to obtain a second neural network layer with random length;
step 411: establishing a second weight matrix Wmaxout2Mapping the second neural network layer by the following formula to obtain a second hidden layer, selecting the neuron with the maximum value in the second hidden layer 2 to map the neuron into a (0,1) interval, and taking the mapped data as the output result of the discriminator;
q’i=w’i1×q1+w’i2×q2+....+w’in×qn
W’j1is a weight matrix Wmaxout2Line j 1 st element, q'1For the 1 st data of the input sequence, qjIs the jth data in the second hidden layer.
In step 411, when the output result of the discriminator is 0.5, the discriminator cannot determine the source of the received data sample; otherwise, the discrimination results in that the received data sample originates from the generator.
In step 5, the constraints of the CGAN model are:
wherein, G (z)noiseY) is sample data generated by the generator according to the condition information Y and the noise information znoise; d (G (z)noiseY)) is the output of the discriminator;replacing with the mean of the generated sample data and the real sample data for log D (x | y) mathematical expectation;for z is a random number, Pz(z) is a probability distribution function of the random number z,obey a probability distribution P when zz(z) mathematical expectation; pdata(x)Probability distribution of input real sample data; p is a probability distribution; v (D, G) represents the game function between the generator and the arbiter.
In said step 5, the generator is optimized by the following formula:
wherein G is the constraint condition of the encoder,for mathematical expectation, Pdata(x | y) is the probability that x is from the true sample data, P, under the conditional constraint of yG(x | y) is the probability that x comes from the sample data generated by the generator under the conditional constraint of y;
when the value of the formula is minimum, the generator is considered to be optimized.
In said step 5, the discriminator is optimized by the following formula:
wherein D is a discriminator constraint condition;
the discriminator optimization is considered to be completed when the formula takes the maximum value.
When the output result of the discriminator is 0.5, the game between the discriminator and the generator is considered to be balanced, the training is finished, and the trained generator is output;
the invention has the following beneficial effects:
the energy storage output is predicted by using the VAE-CGAN model, so that not only can condition information under a historical scene be modeled, but also unknown condition information can be dealt with, the condition information hidden in an input sequence is trained through the VAE model, the CGAN model of a continuous condition input space is constructed, an optimal generator is obtained through confrontation, and the accuracy of energy storage output prediction is guaranteed. Therefore, the optimal output of the stored energy under the condition of large working condition change can be determined more comprehensively and accurately.
Drawings
FIG. 1 is a flow chart of a VAE-CGAN-based energy storage processing prediction method
FIG. 2 VAE training procedure
FIG. 3 CGAN Game model
FIG. 4 is a block diagram of a CGAN neural network
Detailed Description
The following detailed description of the embodiments of the invention is provided in connection with the accompanying drawings.
As shown in fig. 1, the energy storage output prediction method based on VAE-CGAN of the present invention specifically includes the following steps:
step 1: the method comprises the steps of collecting battery voltage, battery ampere hours, discharge rate C and other data such as the power generation ratio of traditional energy and new energy to serve as input sequences of a VAE model, collecting different input sequences to serve as input of the VAE for extracting features according to different systems used by an energy storage battery in actual conditions, wherein the feature sequences with strong correlation such as the battery voltage, the battery ampere hours and the discharge rate C form a multi-element time sequence, analyzing the multi-element time sequence by using the same VAE, and training the data without correlation such as the power generation ratio of the traditional energy and the new energy by using a single VAE. Here, the VAE model input sequence is represented by X ═ X1, X2, … xn ].
Step 2: and (3) training various types of data acquired in the step (1) by using a VAE model to obtain a characteristic information sequence.
Specifically, as shown in fig. 2, step 2 includes the following steps:
step 201: adopting a layer of LSTM (Long short-term memory) neural network to construct an encoder; a decoder is constructed using a layer of LSTM neural network and a layer of fully-connected layers.
Step 202: the encoder receives a first high-dimensional data sequence X ═ X of historical operating data1,x2,…xn]Then, mapping the high-dimensional data sequence X of the historical operating data into a mean vector mu with the length of m and a standard deviation vector sigma with the length of m.
Step 203: the encoder calculates the mean vector mu, the standard deviation vector sigma and the parameter sequence delta [ delta 1, delta 2, … delta m ═ m]And calculating a hidden variable sequence Z ═ Z by the following formula1,z2,…zm]。
zi=μi+δi·exp(σi) (1)
Wherein z isiIs the ith value, mu, in the hidden variable sequence ZiIs the ith value, σ, in the mean vector μiFor the ith value in the standard deviation vector σ, δ i is a parameter obtained by random sampling in a sample set that obeys a bernoulli distribution, and obeys a standard normal distribution, i.e., δ i to N (0,1), i ═ 1,2, …, m.
Step 204: inputting the hidden variable sequence Z obtained by calculation in step 203 into a decoder, restoring the hidden variable sequence Z into a second high-dimensional data sequence X ', wherein X' is [ X1 ', X2', … xn '], that is, the hidden variable sequence Z is identical in dimension to the VAE model input sequence, and finally reconstructing the high-dimensional data feature X' by the full connection layer to obtain a generated m-dimensional feature information sequence Y.
Specifically, the total constraint L in the VAE model training process is calculated and expressed by the following formula;
L=max loss=∫q(z|x)log P(x|z)dz-KL(q(z|x)||P(z)) (2)
the former part of the objective function is the self-coding reconstruction error, which is also called variation lower limit, and the maximum value is taken to represent that when X is used as the input of the encoder, m hidden variables zi are extracted from the encoder in a coding mode, and finally the decoder can recover X from Z with the maximum probability.
The self-encoded reconstruction error L1 is calculated by the following equation:
in the continuous optimization of the encoder, the self-encoding reconstruction error is continuously increased by changing parameters in the LSTM neural network.
The second part is KL divergence, and when the second part takes the minimum value, σi=0,μiWhen z is 0, z can be obtained from the formula (1)iFollowing a standard normal distribution, q (z | x) ═ p (z), KL divergence is zero. The KL divergence is calculated by the following formula:
p (x | Z) is an a priori distribution of hidden variable Z representing the decoder in the VAE, q (Z | x) is a Z a posteriori distribution derived from x representing the encoder in the VAE, it is desirable to fit these two distributions as closely as possible in the VAE model, so P (Z) in the VAE model is a Gaussian distribution with mean 0 and variance 1 when the two fit, KL divergence is 0, when we adjust that q (Z | x) and P (Z | x) are completely consistent, KL divergence vanishes to 0, and Lb and logP (x) are completely consistent. Therefore, no matter what the value of logP (x), we can always make Lb equal to logP (x) by adjusting, and because Lb is the lower bound of logP (x), solving Maximum logP (x) is equivalent to solving Maximum Lb, even if self-coding reconstruction error is Maximum, an optimal encoder model is obtained, and at this time, the trained VAE model can generalize the data characteristics of input data and generate a characteristic sequence as a condition information sequence used in the CGAN model.
And step 3: the method comprises the steps of constructing a generator and a discriminator based on the CGAN, wherein the generator is used for generating simulation sample data, and the discriminator is used for discriminating the sources of the simulation sample data and real sample data.
And 4, step 4: and (3) inputting the characteristic information obtained in the step (2) as condition information into the generator and the discriminator, and performing countermeasure training on the generator and the discriminator, namely, the discriminator receives the simulation sample data and the real sample data acquired in the step (1) and judges the sources of the two data.
And, a random noise information z is set at randomnoiseInput into a generator together with the condition information, the generator utilizes the nonlinear mapping capability of the neural network to give the noise information znoiseAnd condition information Y into simulation sample data G (z)noiseY) is also input to the generator.
In particular, as shown in fig. 3-4, the generation of the immersive sample by the generator comprises in particular the following steps:
step 401: input random noise sequence ZnoiseWhen the length is n and the length of the conditional information sequence Y is m, the generator establishes a full connection layer to enable the generator to be connected with the random noise sequenceThe number of neurons of the full-connection layer corresponding to the condition information sequence Y is respectively equal to the random noise sequenceThe length of the condition information sequence Y is consistent, wherein the full connection layer aggregates and classifies the characteristics of the input sequence under the condition of not changing the length of the input sequence.
Step 402: the data of the fully connected layer is corrected such that its mean value is approximately 0 and its variance is approximately 1, i.e., towards a standard normal distribution N (0, 1). Preferably, in the present invention, the data of the full-link layer is corrected using the Batch-Normalization algorithm.
Step 403: random noise Z to be correctednoiseAnd splicing the correction sequence with the condition information Y correction sequence to form a spliced sequence with the length of (n + m)/10. In one embodiment of the invention, the input random noise sequence ZnoiseIs 200 and the length of the conditional information sequence Y is 1000, a splicing sequence of length 120 is thus obtained in this step.
Step 404: the neurons in the stitched sequence obtained in step 403 are multiplied by the same probability of occurrence p, which is 0.5. Preferably, the concatenation sequence is processed using the dropout algorithm at this step 404.
Step 405: in step 404, about half of the neurons are temporarily deleted in this generation. Therefore, only part of input sequences participate in the generation in the process of generating the sample data by the generator, and the return value only adjusts the neurons participating in the generation, so that the over-fitting phenomenon of the generator can be avoided. The remaining undeleted neurons are output from the generator and input to the discriminator as generation sample data.
Specifically, as shown in fig. 3-4, the step of identifying the source of the data sample by the discriminator specifically comprises the following steps:
step 406: the discriminator receives the condition information Y, the real sample data sequence X and the simulation sample data sequence generated by the generator;
step 407: and the discriminator respectively generates a first hidden layer for the condition information sequence Y, the real sample data sequence X and the simulation sample data sequence, wherein the number of the neurons of the hidden layer is i. Preferably, in this step, the maxout algorithm is adopted to map each data sequence, and specifically, the weight matrix W is setmaxout1And calculating the hidden layer neuron value by the following formula:
t’i=wi1×t1+wi2×t2+....+win×tn (5)
wherein, wilIs a weight matrix Wmaxout1The 1 st element of the ith row, t'iIs the value of the ith neuron in the first hidden layer, tiIs the ith data of the input sequence.
Step 408: dividing neurons in a first hidden layer of a condition information sequence Y, a first hidden layer of a real sample data sequence X and a first hidden layer of a simulation sample data sequence into s groups, selecting the neuron with the largest value from each group of the groups as the output of the first hidden layer, and generating a condition information neural network layer, a real sample data neural network layer and a simulation sample data neural network layer, wherein the neural network layer contains s neurons. Preferably, in the present invention, the neurons in the first hidden layer are divided into 5 groups, thereby obtaining a neural network layer containing 5 neurons.
Step 409: and splicing the generation condition information neural network layer, the real sample data neural network layer and the simulation sample data neural network layer to obtain a first neural network layer, wherein the first neural network layer is provided with 3s neurons. Preferably, a neural network layer containing 15 neurons is obtained in one embodiment of the present invention.
Step 410: setting the neuron deletion probability to be 0.5, and processing the first neural network layer in the step 409 to obtain a second neural network layer with a random length. Specifically, in one embodiment of the present invention, the first neural network layer is processed using the dropout algorithm, and by setting the neuron deletion probability p to 0.5, that is, in one optimization process of the generator model, each of the 15 neurons has a 50% probability of being temporarily deleted until a part of the neurons is restored and deleted again with the probability p in the next optimization process. A generated neural network layer of random length is obtained.
Step 411: and mapping the second neural network layer by the following formula to obtain a second hidden layer, selecting the neuron with the maximum value in the second hidden layer 2 to map into a (0,1) interval, and taking the mapped data as the output result of the discriminator. That is, when the output result of the discriminator is 0.5, the discriminator cannot judge the source of the received data sample; otherwise, the discrimination results in that the received data sample originates from the generator.
q’i=w’i1×q1+w’i2×q2+....+w’in×qn (6)
W’j1Is a weight matrix Wmaxout2Line j 1 st element, q'1For the 1 st data of the input sequence, qjIs the jth data in the second hidden layer.
Preferably, in step 411, the maxout algorithm is used to map the second neural network layer; and mapping the neurons of the second hidden layer by adopting a sigmoid function.
And 5: the parameters of the generator and discriminator are updated, i.e. the generator and discriminator are optimized. Specifically, after the generator completes one generation, the neural network calculates a loss function of the generated data and propagates the loss function back to modify the parameters of the generator. When the generator completes the adjustment of the parameters, the discriminator judges the generated sample sequence, and when the output data of the discriminator is 0.5, the discriminator cannot judge that the data comes from the generator. When the discriminator output is not equal to 0.5, the generator continues the above-described generation and parameter adjustment process until the discriminator cannot distinguish the source of the sequence of samples generated by the generator. The probability that the generator generates the sample sequence at the moment is considered to be derived from real data by the discriminator, the discriminator also has the probability of 50% to judge that the sample sequence generated by the generator comes from the generator, namely the game between the generator and the generator reaches an equilibrium state, and the obtained generator model parameters can be considered as an optimal generator, namely a generator model for energy storage output prediction is established.
The constraint conditions of the CGAN model are as follows:
wherein, G (z)noiseY) is sample data generated by the generator according to the condition information Y and the noise information znoise; d (G (z)noiseY)) is the output of the discriminator;as a mathematical expectation of log D (x | y), the generated sample data and the mean of the real sample data are used instead;for z is a random number, Pz(z) is a probability distribution function of the random number z,obey a probability distribution P when zz(z) mathematical expectation; pdata(x)Probability distribution of input real sample data; pz(z) is a probability distribution function for the random number z; p is a probability distribution; v (D, G) represents the game function between the generator and the arbiter.
In said step 5, the generator and the discriminator are optimized by the following two equations, respectively:
wherein the generator constraint is as follows (7):
wherein the discriminator constraint is the following equation (8):
according to respective constraint conditions, a generator and a discriminator in the CGAN model continuously adjust model parameters to enable a constraint function of the generator to be minimum and model parameters of the discriminator to be maximum, and with the aim of optimizing the generator parameters through back propagation of a neural network, the generator parameters comprise a mapping function of a full connection layer in the generator, values on neurons, the number of the neurons and the like.
Step 6: the number of iterations is increased by 1 and the process returns to step 5 until the discriminator cannot distinguish the source of the sample data. Namely, when the output result of the discriminator is 0.5, the game between the discriminator and the generator is considered to be balanced, the training is finished, so that a stronger generator is obtained, and the generator is output as an output prediction model of the energy storage device.
And 7: and (4) acquiring real-time operation data of the energy storage device, inputting the real-time operation data into the generator output in the step (6), and predicting the energy storage output condition of the energy storage device.
The VAE-CGAN model is used for predicting the energy storage output, not only can condition information under a historical scene be modeled, but also unknown condition information can be responded, the condition information hidden in an input sequence is trained through the VAE model, the continuous condition input space input CGAN model is constructed, an optimal generator is obtained through confrontation, and the accuracy of the energy storage output prediction is guaranteed. Therefore, the optimal output of the stored energy under the working environment with strong uncertainty can be determined more comprehensively and accurately.
While the best mode for carrying out the invention has been described in detail and illustrated in the accompanying drawings, it is to be understood that the same is by way of illustration and example only and is not to be taken by way of limitation, the scope of the invention should be determined by the appended claims and any changes or modifications which fall within the true spirit and scope of the invention should be construed as broadly described herein.
Claims (12)
1. A VAE-CGAN-based energy storage output prediction method is characterized by comprising the following steps:
step 1: collecting historical operation data of the energy storage device, wherein the operation data comprises battery voltage, battery ampere hours, discharge multiplying power, battery operation voltage, traditional energy power generation ratio and new energy power generation ratio;
step 2: directly inputting the historical operating data acquired in the step 1 into a VAE model for training to generate data characteristic information, namely generating new data containing information of the historical operating data;
and step 3: constructing a generator and a discriminator based on the CGAN, wherein the generator is used for generating simulation sample data, and the discriminator is used for discriminating the sources of the simulation sample data and real sample data;
and 4, step 4: inputting the characteristic information obtained in the step 2 as condition information into the generator and the discriminator, and performing countermeasure training on the generator and the discriminator, namely, the discriminator receives the simulation sample data and the real sample data acquired in the step 1 and judges the sources of the two data;
and 5: updating the parameters of the generator and the discriminator, namely optimizing the generator and the discriminator;
step 6: adding 1 to the iteration times and returning to the step 5 until the discriminator cannot distinguish the sample data source, and outputting a trained generator;
and 7: and (4) acquiring real-time operation data of the energy storage device, inputting the real-time operation data into the generator output in the step (6), and predicting the energy storage output condition of the energy storage device.
2. The VAE-CGAN-based energy storage capacity contribution prediction method of claim 1, wherein:
the step 2 comprises the following steps:
step 201: constructing an encoder by adopting a layer of LSTM neural network; a decoder is constructed by adopting a layer of LSTM neural network and a layer of full connection layer;
step 202: the encoder receives a first high-dimensional data sequence X ═ X of historical operating data1,x2,…xn]Then, mapping a high-dimensional data sequence X of the historical operating data into a mean vector mu with the length of m and a standard deviation vector sigma with the length of m;
step 203: the encoder calculates a mean vector mu, a standard deviation vector sigma and a parameter sequence delta [ delta 1, delta 2, … delta m ═ according to the mean vector mu and the standard deviation vector sigma]And calculating a hidden variable sequence Z ═ Z by the following formula1,z2,…zm];
zi=μi+δi·exp(σi)
Wherein z isiFor the ith value in the hidden variable sequence Z,μiIs the ith value, σ, in the mean vector μiIs the ith value in the standard deviation vector σ, δ i is a parameter obtained by random sampling in a sample set subject to bernoulli distribution, and subject to a standard normal distribution, i.e., δ i to N (0,1), i ═ 1,2, …, m; step 204: inputting the hidden variable sequence Z obtained by calculation in the step 203 into a decoder, and restoring the hidden variable sequence Z into a second high-dimensional data sequence X ', wherein the second high-dimensional data sequence X' is a characteristic information sequence.
3. The VAE-CGAN-based energy storage capacity contribution prediction method of claim 2, wherein:
in step 2, the following objective function is adopted for training:
wherein L is1Is a self-encoding reconstruction error, L2Is KL divergence, P (x | Z) is a prior distribution of hidden variables Z representing a decoder in the VAE, q (Z | x) is a posterior distribution of values Z in a sequence of hidden variables derived from any value x in the real data sequence representing an encoder in the VAE;
and when the self-coding reconstruction error value is maximum and the KL divergence value is minimum, finishing training to obtain the characteristic information.
4. The VAE-CGAN based energy storage capacity contribution prediction method of any of claims 1-3, wherein:
in the step 2, the battery voltage, the battery ampere hour and the discharge multiplying power are trained by adopting the same VAE model;
and the traditional energy power generation ratio and the new energy power generation ratio are trained by adopting independent VAE models respectively.
5. The VAE-CGAN based energy storage capacity contribution prediction method of any of claims 1-3, wherein:
in step 4, random noise information Z is setnoiseAnd input into a generator together with the condition information, the generator utilizes the nonlinear mapping capability of the neural network to give the noise information znoiseAnd condition information Y into simulation sample data G (z)noise|y)。
6. The VAE-CGAN-based energy storage capacity contribution prediction method of claim 5, wherein:
in step 4, the generator generating the simulation sample comprises the following steps:
step 401: input random noise sequence ZnoiseWhen the length is n and the length of the conditional information sequence Y is m, the generator establishes a full connection layer to enable the generator to be connected with the random noise sequence ZnoiseThe number of the neurons of the full connection layer corresponding to the condition information sequence Y and the random noise sequence Z are respectivelynoiseThe length of the conditional information sequence Y is consistent with that of the conditional information sequence Y;
step 402: correcting the data of the full-connection layer to enable the mean value to be approximate to 0 and the variance to be approximate to 1, namely, the data are approximate to the standard normal distribution N (0, 1);
step 403: random noise Z to be correctednoiseSplicing the correction sequence with the condition information Y correction sequence to form a spliced sequence with the length of (n + m)/10;
step 404: multiplying the same appearance probability p by (n + m)/10 neurons in the spliced sequence with the length of (n + m)/10 obtained in the previous layer to be 0.5;
step 405: in step 404, half of the neurons are temporarily deleted in the current generation, and the undeleted neurons are output from the generator as the simulation sample data.
7. The VAE-CGAN based energy storage capacity contribution prediction method of any of claims 1-3, wherein: in step 4, the discriminator discriminating the source of the input data comprises the following steps:
step 406: the discriminator receives the condition information Y, the real sample data sequence X and the simulation sample data sequence generated by the generator;
step 407: the discriminator respectively generates a first hidden layer for the condition information sequence Y, the real sample data sequence X and the simulation sample data sequence, wherein the number of the neurons of the hidden layer is i, and a first weight matrix W is establishedmaxout1Calculating the numerical value of the neuron in the first hidden layer by the following formula;
t’i=wi1×t1+wi2×t2+....+win×tn
wherein, wilIs a weight matrix Wmaxout1The 1 st element of the ith row, t'iIs the value of the ith neuron in the first hidden layer, tiIs the ith data of the input sequence;
step 408: dividing neurons in a first hidden layer of a condition information sequence Y, a first hidden layer of a real sample data sequence X and a first hidden layer of a simulation sample data sequence into s groups, selecting the neuron with the largest value from each group of the groups as the output of the first hidden layer, and generating a condition information neural network layer, a real sample data neural network layer and a simulation sample data neural network layer, wherein the neural network layer contains s neurons;
step 409: splicing the generation condition information neural network layer, the real sample data neural network layer and the simulation sample data neural network layer to obtain a first neural network layer, wherein the first neural network layer is provided with 3s neurons;
step 410: setting the neuron deletion probability to be 0.5, and processing the first neural network layer in the step 409 to obtain a second neural network layer with random length;
step 411: establishing a second weight matrix Wmaxout2And layering the second neural network by the following formulaLine mapping is carried out, a second hidden layer is obtained, a neuron with the largest value in the second hidden layer 2 is selected to be mapped into a (0,1) interval, and the mapped data are used as an output result of the discriminator;
q’i=w’i1×q1+w’i2×q2+....+w’in×qn
W’j1is a weight matrix Wmaxout2Line j 1 st element, q'1For the 1 st data of the input sequence, qjIs the jth data in the second hidden layer.
8. The VAE-CGAN-based energy storage capacity contribution prediction method of claim 7, wherein:
in step 411, when the output result of the discriminator is 0.5, the discriminator cannot determine the source of the received data sample; otherwise, the discrimination results in that the received data sample originates from the generator.
9. The VAE-CGAN based energy storage capacity contribution prediction method of any of claims 1-3, wherein:
in step 5, the constraints of the CGAN model are:
wherein, G (z)noiseY) is sample data generated by the generator according to the condition information Y and the noise information znoise; d (G (z)noiseY)) is the output of the discriminator;replacing with the mean of the generated sample data and the real sample data for log D (x | y) mathematical expectation;for z is a random number, Pz(z) is a probability distribution function of the random number z,obey a probability distribution P when zz(z) mathematical expectation; pdata(x)Probability distribution of input real sample data; p is a probability distribution; v (D, G) represents the game function between the generator and the arbiter.
10. The VAE-CGAN-based energy storage capacity contribution prediction method of claim 9, wherein:
in said step 5, the generator is optimized by the following formula:
wherein G is the constraint condition of the encoder,for mathematical expectation, Pdata(x | y) is the probability that x is from the true sample data, P, under the conditional constraint of yG(x | y) is the probability that x comes from the sample data generated by the generator under the conditional constraint of y;
when the value of the formula is minimum, the generator is considered to be optimized.
11. The VAE-CGAN-based energy storage capacity contribution prediction method of claim 9, wherein:
in said step 5, the discriminator is optimized by the following formula:
wherein D is a discriminator constraint condition;
the discriminator optimization is considered to be completed when the formula takes the maximum value.
12. The VAE-CGAN based energy storage capacity contribution prediction method of any of claims 1-3, wherein:
in step 6, when the result of the discriminator output is 0.5, the game between the discriminator and the generator is considered to be balanced, and the training is finished.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN202011315545.3A CN112508239A (en) | 2020-11-22 | 2020-11-22 | Energy storage output prediction method based on VAE-CGAN |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN202011315545.3A CN112508239A (en) | 2020-11-22 | 2020-11-22 | Energy storage output prediction method based on VAE-CGAN |
Publications (1)
Publication Number | Publication Date |
---|---|
CN112508239A true CN112508239A (en) | 2021-03-16 |
Family
ID=74959347
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN202011315545.3A Pending CN112508239A (en) | 2020-11-22 | 2020-11-22 | Energy storage output prediction method based on VAE-CGAN |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN112508239A (en) |
Cited By (1)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN112950409A (en) * | 2021-04-19 | 2021-06-11 | 工数科技(广州)有限公司 | Production scheduling optimization method of gas and steam energy comprehensive utilization system |
Citations (10)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US20180082172A1 (en) * | 2015-03-12 | 2018-03-22 | William Marsh Rice University | Automated Compilation of Probabilistic Task Description into Executable Neural Network Specification |
CN109859310A (en) * | 2019-01-22 | 2019-06-07 | 武汉纺织大学 | A kind of model and its method for building up can be used for generating MR image |
CN109886970A (en) * | 2019-01-18 | 2019-06-14 | 南京航空航天大学 | The detection dividing method and computer storage medium of target object in terahertz image |
CN110245380A (en) * | 2019-05-10 | 2019-09-17 | 西安理工大学 | Soft instrument training and sample compensation process |
GB201911689D0 (en) * | 2019-08-15 | 2019-10-02 | Facesoft Ltd | Facial image processing |
US20200074269A1 (en) * | 2018-09-05 | 2020-03-05 | Sartorius Stedim Data Analytics Ab | Computer-implemented method, computer program product and system for data analysis |
CN111037365A (en) * | 2019-12-26 | 2020-04-21 | 大连理工大学 | Cutter state monitoring data set enhancing method based on generative countermeasure network |
CN111191835A (en) * | 2019-12-27 | 2020-05-22 | 国网辽宁省电力有限公司阜新供电公司 | IES incomplete data load prediction method and system based on C-GAN transfer learning |
CN111275115A (en) * | 2020-01-20 | 2020-06-12 | 星汉智能科技股份有限公司 | Method for generating counterattack sample based on generation counternetwork |
CN111598805A (en) * | 2020-05-13 | 2020-08-28 | 华中科技大学 | Confrontation sample defense method and system based on VAE-GAN |
-
2020
- 2020-11-22 CN CN202011315545.3A patent/CN112508239A/en active Pending
Patent Citations (10)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US20180082172A1 (en) * | 2015-03-12 | 2018-03-22 | William Marsh Rice University | Automated Compilation of Probabilistic Task Description into Executable Neural Network Specification |
US20200074269A1 (en) * | 2018-09-05 | 2020-03-05 | Sartorius Stedim Data Analytics Ab | Computer-implemented method, computer program product and system for data analysis |
CN109886970A (en) * | 2019-01-18 | 2019-06-14 | 南京航空航天大学 | The detection dividing method and computer storage medium of target object in terahertz image |
CN109859310A (en) * | 2019-01-22 | 2019-06-07 | 武汉纺织大学 | A kind of model and its method for building up can be used for generating MR image |
CN110245380A (en) * | 2019-05-10 | 2019-09-17 | 西安理工大学 | Soft instrument training and sample compensation process |
GB201911689D0 (en) * | 2019-08-15 | 2019-10-02 | Facesoft Ltd | Facial image processing |
CN111037365A (en) * | 2019-12-26 | 2020-04-21 | 大连理工大学 | Cutter state monitoring data set enhancing method based on generative countermeasure network |
CN111191835A (en) * | 2019-12-27 | 2020-05-22 | 国网辽宁省电力有限公司阜新供电公司 | IES incomplete data load prediction method and system based on C-GAN transfer learning |
CN111275115A (en) * | 2020-01-20 | 2020-06-12 | 星汉智能科技股份有限公司 | Method for generating counterattack sample based on generation counternetwork |
CN111598805A (en) * | 2020-05-13 | 2020-08-28 | 华中科技大学 | Confrontation sample defense method and system based on VAE-GAN |
Non-Patent Citations (6)
Title |
---|
IAN J. GOODFELLOW 等: "Generative Adversarial Nets", 《ADVANCES IN NEURAL INFORMATION PROCESSING SYSTEMS》 * |
PRABHAT 等: "Comparative Analysis of Deep Convolutional Generative Adversarial Network and Conditional Generative Adversarial Network using Hand Written Digits", 《2020 4TH INTERNATIONAL CONFERENCE ON INTELLIGENT COMPUTING AND CONTROL SYSTEMS (ICICCS)》 * |
YIZE CHEN 等: "Model-Free Renewable Scenario Generation Using Generative Adversarial Networks", 《IEEE TRANSACTIONS ON POWER SYSTEMS》 * |
张文强 等: "基于VAE-CGAN的光伏不确定性建模方法", 《电网技术》 * |
杨英 等: "VAE_LSTM算法在时间序列预测模型中的研究", 《湖南科技大学学报》 * |
梁俊杰 等: "生成对抗网络GAN综述", 《计算机科学与探索》 * |
Cited By (1)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN112950409A (en) * | 2021-04-19 | 2021-06-11 | 工数科技(广州)有限公司 | Production scheduling optimization method of gas and steam energy comprehensive utilization system |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN110059844B (en) | Energy storage device control method based on ensemble empirical mode decomposition and LSTM | |
CN110705743B (en) | New energy consumption electric quantity prediction method based on long-term and short-term memory neural network | |
CN110212528B (en) | Power distribution network measurement data missing reconstruction method | |
CN110309603A (en) | A kind of short-term wind speed forecasting method and system based on wind speed characteristics | |
CN110648017A (en) | Short-term impact load prediction method based on two-layer decomposition technology | |
CN111144644B (en) | Short-term wind speed prediction method based on variation variance Gaussian process regression | |
CN113553768B (en) | Method and device for rapidly calculating reliability of power grid driven by model data in hybrid mode | |
CN110837915B (en) | Low-voltage load point prediction and probability prediction method for power system based on hybrid integrated deep learning | |
CN112884236B (en) | Short-term load prediction method and system based on VDM decomposition and LSTM improvement | |
CN112834927A (en) | Lithium battery residual life prediction method, system, device and medium | |
CN113705086A (en) | Ultra-short-term wind power prediction method based on Elman error correction | |
CN114971090A (en) | Electric heating load prediction method, system, equipment and medium | |
CN109146131A (en) | A kind of wind-power electricity generation prediction technique a few days ago | |
CN112149896A (en) | Attention mechanism-based mechanical equipment multi-working-condition fault prediction method | |
CN112116171A (en) | Novel photovoltaic power generation power prediction method based on neural network | |
CN111697560A (en) | Method and system for predicting load of power system based on LSTM | |
CN113762591B (en) | Short-term electric quantity prediction method and system based on GRU and multi-core SVM countermeasure learning | |
CN112508239A (en) | Energy storage output prediction method based on VAE-CGAN | |
CN108537581B (en) | Energy consumption time series prediction method and device based on GMDH selective combination | |
CN112232570A (en) | Forward active total electric quantity prediction method and device and readable storage medium | |
CN113919594A (en) | Demand response potential evaluation method based on deep forest | |
CN113435595A (en) | Two-stage optimization method for extreme learning machine network parameters based on natural evolution strategy | |
CN117291069A (en) | LSTM sewage water quality prediction method based on improved DE and attention mechanism | |
Xu et al. | Short-term electricity consumption forecasting method for residential users based on cluster classification and backpropagation neural network | |
CN111724277A (en) | New energy and multi-element load value matching 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 |