CN110429652B - Intelligent power generation control method capable of expanding deep width self-adaptive dynamic planning - Google Patents
Intelligent power generation control method capable of expanding deep width self-adaptive dynamic planning Download PDFInfo
- Publication number
- CN110429652B CN110429652B CN201910801884.3A CN201910801884A CN110429652B CN 110429652 B CN110429652 B CN 110429652B CN 201910801884 A CN201910801884 A CN 201910801884A CN 110429652 B CN110429652 B CN 110429652B
- Authority
- CN
- China
- Prior art keywords
- network
- width
- deep
- tri
- power generation
- 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
Links
Images
Abstract
The invention provides an intelligent power generation control method capable of expanding a deep width self-adaptive dynamic programming, which combines an expandable deep width self-adaptive dynamic programming algorithm with a learning rate self-adaptive algorithm and is suitable for active control of 'source-load-bearing' tri-state energy under the unified time scale of a Future Integrated Energy Interconnection System (FIEIS). The method solves the problem that the traditional power generation control is difficult to adapt to real-time flexible control of active frequency under the high permeability of the tri-state energy and the intermittent energy of the power system, has strong robustness and stability, and improves the iterative convergence speed and accuracy. According to the method, the depth and the width of the deep neural network are adaptively changed according to the number of the tri-state energy types of the access network, so that the real-time performance, the accuracy and the stability of active control are enhanced. The depth-width model prediction network, the evaluation network and the execution network of the method effectively replace the traditional multi-time scale combined active control algorithm.
Description
Technical Field
The invention relates to an intelligent power generation control method capable of expanding deep width self-adaptive dynamic programming, belongs to the field of intelligent power generation scheduling and control of a power system, and is applied to tri-state energy active frequency control of 'source load' under unified time scale of a future comprehensive energy interconnection system.
Background
With the increase of the permeability of intermittent new energy sources such as wind power generation and photovoltaic power generation, the economic optimization configuration and stable regulation of the active frequency of a power grid face challenges. In order to ensure safe and economic operation of a power grid, the problem of difference of output active frequency is solved, active power is adjusted in the traditional power generation modes of hydropower, thermal power and the like, and tri-state energy sources with energy storage functions, such as electric vehicles, power batteries, flywheel energy storage and the like, can be connected into the power grid to partially balance system load active power and frequency fluctuation in a charging and discharging mode.
When an electric automobile, a power battery and the like are introduced to participate in active frequency regulation of a system, the generated active power has the characteristics of random uncertainty and source-load, which increases the difficulty for regulation and control of a power grid. Taking the active power regulation of the electric vehicle participation system as an example, the active power random uncertainty is realized when the use of the electric vehicle is influenced by multiple factors such as user habits, weather conditions, holiday travel and the like; the "source charge" of the electric vehicle means that the electric vehicle can be used as a power generation end state, a power utilization end state and a shutdown state, namely a tri-state energy source, and the state of the electric vehicle is changed in real time. Besides the influence of the tri-state energy on the power grid active power, the large-scale access of intermittent new energy such as wind power, tidal power generation and photovoltaic power generation can also increase the load of power grid electric energy active power control.
In order to solve the real-time regulation and control of the active frequency of a distributed new energy access power grid, particularly to meet the higher requirement of industrial upgrading on the electric energy quality, the construction of a smart power grid in the future must be capable of accommodating more tri-state energy. The smart grid should also be a robust grid with both economy, reliability and flexibility. In the field of algorithms for implementing optimal control of power systems, expert scholars have proposed a number of solutions. For example, a power frequency distribution method solved by deep learning and an intelligent power generation method of adaptive width dynamic planning solve active regulation of the system to a certain extent. However, in the face of increasingly complex power system environments, especially in the high-permeability complex environments of future tri-state energy sources and distributed new energy sources, the problem of lag is caused by the control in the traditional multi-time scale method. Therefore, an algorithm with strong learning ability and better robustness is still required to solve the problem of tri-state energy comprehensive active control of the unified time scale of the Future Integrated Energy Interconnection System (FIEIS).
Disclosure of Invention
The invention provides an intelligent power generation control method capable of expanding deep width self-adaptive dynamic programming, which aims to solve the problem that high-quality control of active power and frequency is difficult to realize in a traditional combined power generation control mode after FIEIS grid-connected power generation. According to the traditional combined power generation dispatching control algorithm based on multi-time scale dispatching control, after the tri-state energy scale grid connection, frequency deviation caused by power random fluctuation cannot be regulated and controlled in time. The automatic power generation control of the unified time scale takes 4s as a period, so that the unit combination and economic dispatching distribution delay caused by the original multi-time scale dispatching is avoided.
The control method based on the expandable depth-width self-adaptive dynamic programming provided by the invention is different from the neural network of the traditional self-Adaptive Dynamic Programming (ADP) algorithm on the frame for solving the problem of adjusting and controlling the active power frequency of power generation. But rather a deep neural network based depth and width scalable algorithmic framework, whose networks are stacked with several constrained boltzmann machines (RBMs). The restricted Boltzmann machine comprises a visible layer and a hidden layer, and no connection exists between units in the same layer. The energy function of the limited boltzmann machine is as follows:
in the formula riFor the biasing of the visible layer cell i, cjTo hide the bias of layer element j, LijConnecting weights for cell i and cell j, ρ being the parameter set of all connection weights and offsets, viIs a visible layer unit i, hjTo hide layer element j, with vi×hjDenotes viAnd hjThe degree of association of (c). The joint probability distribution of the state variables (v, h) is expressed as:
f(v,h|ρ)=exp{-E(v,h|ρ)}/z (2)
The likelihood function is required to be used in the training process of the limited Boltzmann machine, and the logarithm of the likelihood function is taken for convenient calculation:
the above equation is biased to calculate the gradient:
the cell offset r is then updatedi、cjAnd a connection weight LijAnd (3) equal parameters:
the offset derivative in the above equation is expressed as:
after the parameter calculation of each layer is completed, the activation condition of the visible layer unit is solved, and the activation probability represents that:
in the formula, sigma is an activation function, in order to avoid the problem of gradient disappearance of training, a ReLU activation function can be used in the deep neural network to solve the problem of nonlinear classification and learning, and the expression isThe activation probability of the hidden layer unit is then expressed as:
the probability of an action translates to:
in the formula kaRepresenting the action probability coefficient, the state of the probability transitions to:
in the formula tsThe probability transformation coefficients are represented.
The intelligent power generation control method based on the self-adaptive dynamic programming capable of expanding the depth width has no network prediction capability in the initial state, and needs to be trained offline in advance for accurately obtaining the system operation characteristics. And (4) performing offline training by adopting an unsupervised greed layer-by-layer training mode, training past calibrated data samples and obtaining initial values of network weights. In order to ensure the learning correctness of the algorithm, the samples are input and output data which can comprehensively reflect the real running state of the past system. Second, to prevent the samples from being over-fitted, techniques such as Dropout can be used to improve the system network generalization capability. When the on-line control system generates power, real-time data is trained in a supervised training mode, and off-line training results are finely adjusted, so that learning results are continuously updated.
The self-adaptive dynamic programming power generation method capable of expanding the deep width has the width and the depth expansion depending on the number of tri-state energy sources and distributed energy sources accessed to a power system network. As previously mentioned, the tri-state energy source of the access system exists in power generation, power usage and shutdown states. And adjusting the depth and the width of the deep neural network on line according to different states of the tri-state energy and the starting and stopping states of the distributed energy of the access system so as to improve the accuracy of obtaining system characteristics and predicting system operating parameters.
Drawings
FIG. 1 is a schematic diagram of an extensible depth and width neural network of the method of the present invention.
Fig. 2 is a schematic diagram of the intermittent new energy source and the tri-state energy source power generation curve of the method.
FIG. 3 is a schematic diagram of a framework of a unified time scale power generation control method applied by the method of the present invention.
Detailed Description
The invention provides an intelligent power generation control method capable of expanding deep width self-adaptive dynamic planning, which is described in detail in combination with the accompanying drawings as follows:
FIG. 1 is a schematic diagram of an extensible depth and width neural network of the method of the present invention. The deep width neural network comprises a deep width model prediction network, a deep width evaluation network and a deep width execution network. These networks are similar in structure, differing only in the input and output content of the network. The expanded deep-width neural network in the figure comprises an input layer, an output layer and 2 hidden layers. In the figure, the number mi,m0The number of neural units representing the input layer and the output layer; in m1、m2、…、mjIndicating the number of individual hidden layer elements. When the number of the tri-state energy or the distributed energy of the access system changes, the depth model predicts the network to adapt to the access action set and the state number, the width is expanded by increasing the number of the input layer units, and the network is changed from the original { m }i,m1,m2,m0Broadening to mi+Δmi,2(m1+Δmi),m2,m0}; the deep execution network can expand the number of output layer units, and the network is changed from the original { mi,m1,m2,m0Broadening to mi,m1,2(m2+Δmo),mo+ΔmoAnd increasing the output action quantity to be fed back to the access topology node to control the active power and the frequency deviation of the power system. In order to adapt to the increase of learning complexity and the self-adaptive expansion of network depth of the deep-width neural network, the original network can be changed into { m }i,m1,m2,m3,m4,m5,moOr both the depth and width of the network.
Fig. 2 is a schematic diagram of the intermittent new energy source and the tri-state energy source power generation curve of the method. In the figure, the energy sources such as wind power generation, photovoltaic power generation and tidal power generation have intermittent characteristics, the quality of different power generation types is different, and the penetration rate of intermittent new energy sources in a power grid can be further improved in the future. As shown in the figure, the tri-state energy source has three states, power, load and shutdown. After the electric automobile, the energy storage battery and the pumped storage power generation are connected into a power grid, the active power and the frequency of a balance system can be adjusted by utilizing the balance system, and the peak clipping and valley filling effects are achieved. Whether intermittent new energy or tri-state energy, the scheduling control of the power system becomes complicated after large-scale access, particularly the power system has intermittent parts, and the power system is difficult to regulate in real time in a traditional power generation control mode. The development of the future smart grid requires stable adjustment of frequency and active power after various forms of power generation energy are accessed.
FIG. 3 is a schematic diagram of a framework of a unified time scale power generation control method applied by the method of the present invention. After the permeability of the tri-state energy and the distributed new energy of the future power grid is improved, the problem of unified regulation and control of local or regional power generation is considered in the framework. The deep-width neural network is trained offline layer by layer in advance, and the deep cognition contains the characteristics of each topological node of the tri-state energy power grid. For a special scene with a large change of the network topology structure and the power supply composition during isolated network operation, pre-training of an intelligent agent needs to be strengthened so as to realize more intelligence of power grid control and more stable system. When the method is used for the on-line training of an actual system, the characteristic data of the power generation and utilization state, the power, the frequency and the like extracted from the topological nodes of the system network are used as an input state parameter matrix S ═ S of the neural network with the expandable deep width after being transformed1,S2,S3,…,Sm]. WhereinAnd k state parameters which represent the generating characteristics of the unit are input by the ith unit (comprising the tri-state energy). Outputting an action matrix A ═ A from a deep-width execution network via internal computation of the agent1,A2,A3,…,Am]WhereinRepresenting k commands generated by the ith unit (including the tri-state energy source). A movesMaking a matrix as a control instruction, adjusting the power generation power and the frequency deviation delta f of the power supply combination and controlling the state of the tri-state energy source; meanwhile, the A matrix is used as a sample and provided for the intelligent agent to train on line, and the control result is output and fine-tuned at the next moment. When the number of power supplies and three-state energy types of the access network is increased, the width and the depth of the deep-width neural network are automatically expanded to adapt to a more complex network.
Claims (1)
1. An intelligent power generation control method capable of expanding deep width self-adaptive dynamic planning is characterized in that the method is used for solving the problem that high-quality control of active power and frequency is difficult to realize in a traditional combined power generation control mode after future comprehensive energy interconnection system grid-connected power generation;
the method has no network prediction capability in an initial state, and needs to be trained offline in advance for accurately obtaining the system operation characteristics; performing offline training by adopting an unsupervised greedy layer-by-layer training mode, training past calibrated data samples and obtaining initial values of network weights; the width and depth of the method are expanded depending on the number of tri-state energy sources and distributed energy sources accessed to the power system network; the deep width neural network of the method comprises a deep width model prediction network, a deep width evaluation network and a deep width execution network; the expanded deep-width neural network of the method comprises an input layer, an output layer and 2 hidden layers; in mi,m0The number of neural units representing the input layer and the output layer; in m1、m2、…、mjRepresenting the number of each hidden layer unit; when the number of the tri-state energy or the distributed energy of the access system changes, the depth model predicts the network to adapt to the access action set and the state number, the width is expanded by increasing the number of the input layer units, and the network is changed from the original { m }i,m1,m2,m0Broadening to mi+Δmi,2(m1+Δmi),m2,m0}; the deep execution network can expand the number of output layer units, and the network is changed from the original { mi,m1,m2,m0Broadening to mi,m1,2(m2+Δmo),mo+Δmo};
When the method is used for on-line training in an actual system, characteristic data of power generation and utilization state, power, frequency and the like extracted from a system network topology node are used as an input state parameter matrix S ═ S of an expandable deep width neural network after being converted1,S2,S3,…,Sm](ii) a WhereinRepresenting k state parameters which can represent the generating characteristics of the unit and are input by the ith unit comprising the tri-state energy; outputting an action matrix A ═ A from a deep-width execution network via internal computation of the agent1,A2,A3,…,Am]WhereinRepresenting k commands generated by the ith unit including the tri-state energy source; the action matrix A is used as a control instruction, the power generation power and the frequency deviation delta f of the power supply combination are adjusted, and the state of the tri-state energy source is controlled; meanwhile, the matrix A is used as a sample and provided for the intelligent agent to train on line, and the control result is output and fine-tuned at the next moment; when the number of power supplies and three-state energy types of the access network is increased, the width and the depth of the deep-width neural network are automatically expanded to adapt to a more complex network;
the tri-state energy accessed into the system has power generation, power utilization and shutdown states, namely the tri-state energy has three states of power supply, load and shutdown; and adjusting the depth and the width of the deep neural network on line according to different states of the tri-state energy and the starting and stopping states of the distributed energy of the access system.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201910801884.3A CN110429652B (en) | 2019-08-28 | 2019-08-28 | Intelligent power generation control method capable of expanding deep width self-adaptive dynamic planning |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201910801884.3A CN110429652B (en) | 2019-08-28 | 2019-08-28 | Intelligent power generation control method capable of expanding deep width self-adaptive dynamic planning |
Publications (2)
Publication Number | Publication Date |
---|---|
CN110429652A CN110429652A (en) | 2019-11-08 |
CN110429652B true CN110429652B (en) | 2021-03-26 |
Family
ID=68417855
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201910801884.3A Active CN110429652B (en) | 2019-08-28 | 2019-08-28 | Intelligent power generation control method capable of expanding deep width self-adaptive dynamic planning |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN110429652B (en) |
Families Citing this family (6)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN110889780B (en) * | 2019-12-04 | 2023-03-31 | 浙江大学 | Wind turbine generator system performance-guaranteeing maximum wind energy capturing method based on BLS and double-layer SMC |
CN111555297B (en) * | 2020-05-21 | 2022-04-29 | 广西大学 | Unified time scale voltage control method with tri-state energy unit |
CN111799820B (en) * | 2020-05-27 | 2022-07-05 | 广西大学 | Double-layer intelligent hybrid zero-star cloud energy storage countermeasure regulation and control method for power system |
CN111769547B (en) * | 2020-06-12 | 2023-07-28 | 广西大学 | Real-time regulation and control method for interactive comprehensive energy system of three-layer linkage mechanism |
CN112564118B (en) * | 2020-11-23 | 2022-03-18 | 广西大学 | Distributed real-time voltage control method capable of expanding quantum deep width learning |
CN115690897B (en) * | 2022-08-31 | 2023-10-20 | 北京夕阳无忧科技有限公司 | Unexpected behavior processing method, device and equipment for preventing privacy leakage and storage medium |
Family Cites Families (4)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN105790314B (en) * | 2016-03-08 | 2017-03-29 | 南京邮电大学 | A kind of distributed generator control method for coordinating based on self adaptation dynamic programming |
CN105846461B (en) * | 2016-04-28 | 2022-01-28 | 中国电力科学研究院 | Control method and system for large-scale energy storage power station self-adaptive dynamic planning |
CN109149648A (en) * | 2018-10-11 | 2019-01-04 | 广西大学 | A kind of adaptive width Dynamic Programming intelligent power generation control method |
CN109193641A (en) * | 2018-10-11 | 2019-01-11 | 广西大学 | A kind of tri-state energy control method based on automatic expansion deep learning |
-
2019
- 2019-08-28 CN CN201910801884.3A patent/CN110429652B/en active Active
Also Published As
Publication number | Publication date |
---|---|
CN110429652A (en) | 2019-11-08 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN110429652B (en) | Intelligent power generation control method capable of expanding deep width self-adaptive dynamic planning | |
Li et al. | Coordinated load frequency control of multi-area integrated energy system using multi-agent deep reinforcement learning | |
CN111884213B (en) | Power distribution network voltage adjusting method based on deep reinforcement learning algorithm | |
CN112186743B (en) | Dynamic power system economic dispatching method based on deep reinforcement learning | |
CN110929948B (en) | Fully distributed intelligent power grid economic dispatching method based on deep reinforcement learning | |
CN111934335A (en) | Cluster electric vehicle charging behavior optimization method based on deep reinforcement learning | |
Aravindh et al. | Design of observer-based non-fragile load frequency control for power systems with electric vehicles | |
CN109494721B (en) | Distributed self-adaptive control method suitable for power distribution network with flexible multi-state switch | |
Khalid et al. | Efficient load frequency control of renewable integrated power system: A twin delayed DDPG-based deep reinforcement learning approach | |
CN106532691A (en) | Adaptive dynamic programming-based frequency compound control method of single-region power system | |
CN110165714A (en) | Micro-capacitance sensor integration scheduling and control method, computer readable storage medium based on limit dynamic programming algorithm | |
CN111891111B (en) | Hybrid electric vehicle interval II type fuzzy logic self-adaptive control method based on MCPSO | |
Yin et al. | Expandable deep width learning for voltage control of three-state energy model based smart grids containing flexible energy sources | |
CN115795992A (en) | Park energy Internet online scheduling method based on virtual deduction of operation situation | |
CN117039981A (en) | Large-scale power grid optimal scheduling method, device and storage medium for new energy | |
Paterova et al. | Estimating harvestable solar energy from atmospheric pressure using deep learning | |
CN114430165A (en) | Micro-grid group intelligent coordination control method and device based on depth model prediction | |
Wu et al. | A dual-driven predictive control for photovoltaic-diesel microgrid secondary frequency regulation | |
CN113394770A (en) | Interconnected micro-grid group frequency complete distribution type optimization control method and system | |
CN114400675B (en) | Active power distribution network voltage control method based on weight mean value deep double-Q network | |
US11407327B1 (en) | Controlling ongoing usage of a battery cell having one or more internal supercapacitors and an internal battery | |
CN112084680A (en) | Energy Internet optimization strategy method based on DQN algorithm | |
Skiparev et al. | Reinforcement learning based MIMO controller for virtual inertia control in isolated microgrids | |
Khooban et al. | Modeling and HiL real-time simulation for the secondary LFC in time-delay shipboard microgrids | |
CN112134304B (en) | Micro-grid full-automatic navigation method, system and device based on deep learning |
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 |