首页 关于本刊 编 委 会 期刊动态 作者中心 审者中心 读者中心 下载中心 联系我们 English
 自动化学报  2018, Vol. 44 Issue (5): 901-914 PDF

1. 东北大学信息科学与工程学院 沈阳 110819

We-energy Hybrid Modeling and Parameter Identification With GAN Technology
SUN Qiu-Ye1, HU Jing-Wei1, YANG Ling-Xiao1, ZHANG Hua-Guang1
1. College of Information Science and Engineering, Northeastern University, Shenyang 110819
Manuscript received : August 31, 2017, accepted: March 7, 2018.
Foundation Item: Supported by the Key Program of National Natural Science Foundation of China (61433004), National Natural Science Foundation of China (61573094), and The Central University Based Research Fees (N140402001)
Corresponding author. SUN Qiu-Ye  Professor at the School of Information Science and Engineering, Northeastern University. His research interest covers network control technology, distributed control technology, distributed optimization analysis and various applications in energy internet, microgrid, power distribution network. Corresponding author of this paper.
Recommended by Associate Editor TAN Ying
Abstract: As a sub-unit of the energy internet, we-energy (WE) aims at realizing bi-directional power transformation and flexible conversion between various types of energies. As the operating characteristics of WE have large difference under different working conditions, existing methods can not accurately identify its parameters. In order to solve this problem, a data-mechanism hybrid driving method based on generative adversarial networks (GAN) is proposed. In order to switch the WE model under different operating conditions, fuzzy theory is used to achieve fuzzy classification of training data of the GAN model by expertise. A modified GAN model containing policy gradient feedback is applied in training model, therefore solving the issue of discrete output sequence of WE. Simulation results validate that the proposed model is of high identification accuracy and has better generalization performance, and can effectively fit the state variation of each node of the whole system under different operation modes.
Key words: Energy internet     we-energy (WE)     generative adversarial networks (GAN)     generative model     adversarial learning     zero-sum game

1 自能源

 图 1 自能源结构 Figure 1 Structure of we-energy
2 自能源机理模型 2.1 电力子系统模型

 图 2 电力子系统模型 Figure 2 Power subsystem model for we-energy

 $$$\tilde {\pmb S}_{DG} = - \left( P_{DG} + {\rm j}Q_{DG} \right)$$$ (1)

 $$$\begin{split} {P_{DG}} =&{P_{DG, 0}} - \frac{1}{m}\left( {f - {f_0}} \right)\\ {Q_{DG}} =&{Q_{DG, 0}} - \frac{1}{n}\left( {{U_{DG}} - {U_0}} \right) \end{split}$$$ (2)

 $$${P_{Q, o}} = {P_{Q, i}} + {P_{Q, EB}} + {P_{Q, MT}} + \Delta {P_{Q, \rm {pump}}} - {P_{Q, L}}$$$ (18)
2.3 天然气网子系统模型

 图 4 天然气子系统模型 Figure 4 Natural gas pipeline model for we-energy

 $$$\Delta {p_{g, 1}} = {\rho _{g, 1}}g{H_g}$$$ (19)

 $$${p_{g, 1}}\Delta {v_g} + {v_{g, 1}}\Delta {p_g} + \Delta {p_g}\Delta {v_g} = {\gamma _1}{P_{\rm {comp}}}$$$ (20)

 $$${P_{g, \rm {tot}}} = {P_M} + {P_Q}$$$ (21)

 $$$\Delta {P_{g, \rm {tot}}} = \Delta {p_{g, 1}} \cdot S_{g, \rm {pipe}} \cdot {v_{g, 1}}$$$ (22)

 $$$\begin{split} \left( {{p_{g, 1}} + \Delta {p_{g, 1}}} \right){v_{g, 1}}S_{g, \rm {pipe}} &+ c{\dot m_{g, 1}}{T_{g, 1}} = \\ {p_{g, 2}}{v_{g, 2}}S_{g, \rm {pipe}} &+ c{\dot m_{g, 2}}{T_{g, 2}} \end{split}$$$ (23)

 $$$\displaystyle {\rho _g}{T_g} = \frac{M}{R}{p_g}$$$ (24)

 $$$\displaystyle {\Delta {p_g}\Delta {v_g}} = {\gamma _1}{P_{\rm {comp}}} + 2{p_{g, 1}}{v_{g, 1}} - {p_{g, 1}}{v_g} - {v_{g, 1}}{p_g}$$$ (25)

 \begin{align} \displaystyle {V_{g, MT}} &= {S_{g, MT}}{v_{g, MT}}\nonumber\\ {V_{g, L}} &= {S_{g, L}}{v_{g, L}} \end{align} (26)

 \begin{align} {\dot m_{g, 2}}&\left( {\frac{{{p_{g, 2}}}}{{{\rho _{g, 2}}g}} + \frac{{v_{g, 2}^2}}{{2g}}} \right) = {\dot m_{g, s}}\left( {\frac{{{p_{g, s}}}}{{{\rho _{g, s}}g}} + \frac{{v_{g, s}^2}}{{2g}}} \right)-\nonumber\\ & {\dot m_{g, MT}} \left( {\frac{{{p_{g, MT}}}} {{{\rho _{g, MT}}g}} + \frac{{v_{g, MT}^2}}{{2g}}} \right)- \nonumber\\ & {\dot m_{g, L}}\left( {\frac{{{p_{g, L}}}}{{{\rho _{g, L}}g}} + \frac{{v_{g, L}^2}}{{2g}}} \right) - {\dot m_{g, L}}\Delta {H_{g, l}} \end{align} (27)

 $$${p_{g, 2}}{\dot V_{g, 2}} = {p_{g, s}}{\dot V_{g, s}} - {p_{g, MT}}{\dot V_{g, MT}} - {p_{g, L}}{\dot V_{g, L}}$$$ (28)

 $$${Z_g} = \frac{{{p_g}{{\dot V}_g}}}{{1 \times {{10}^5}}}$$$ (29)

 $$${Z_{g, 2}} = {Z_{g, 1}} - {Z_{g, MT}} - {Z_{g, L}} + \Delta {Z_{g, \rm {comp}}}$$$ (30)
2.4 自能源整体模型

 \begin{align} \displaystyle {P_E} =&{\theta _{11}}{U^2} + {\theta _{12}}U + {\theta _{13}}f - {\theta _{14}}{v_{g, MT}} + {\theta _{15}}\nonumber\\ {Q_E} =&{\theta _{21}}{U^2} + {\theta _{22}}U + {\theta _{23}}{U_{DG}} + {\theta _{24}}\nonumber\\ {P_Q} =&{\theta _{31}}{v_w}{T_w} - {\theta _{32}}{T_w} + {\theta _{33}}{v_{g, MT}} -\nonumber\\ &{\theta _{34}}{F_{Q, L}} + {\theta _{35}} + {\eta _{EB}}{P_{EB}}\nonumber\\ {Z_g} =&{\theta _{41}}{v_g} + {\theta _{42}}{p_g} + {\theta _{43}}{v_{g, MT}} +\nonumber\\ &{\theta _{44}}{v_{g, L}} + {\theta _{45}}{P_{\rm {comp}}} + {\theta _{46}} \end{align} (31)

3 自能源模型参数辨识的GAN算法实现

 图 5 基于模糊分类的GAN模型 Figure 5 GAN structure based on fuzzy classification
3.1 自能源的模糊分类

 $$$\displaystyle {\mu _{E, U}}\left( U \right) = \left\{ \begin{array}{l} \displaystyle {\frac {{U - U_{\max }^{ur}}}{{U_{\max }^{nor} - U_{\max }^{ur}}}, \quad U_{\max }^{nor} \le U \le U_{\max }^{ur}}\\ 1, \quad\quad\quad\qquad\quad U_{\min }^{nor} \le U \le U_{\max }^{nor}\\ \displaystyle {\frac {{U - U_{\min }^{ur}}}{{U_{\min }^{nor} - U_{\min }^{ur}}}}, \quad U_{\min }^{ur} \le U \le U_{\min }^{nor} \end{array} \right.$$$ (32)
 $$$\displaystyle {\mu _{E, f}}\left( f \right) = \left\{ \begin{array}{l} \displaystyle\frac{{f - f_{\max }^{ur}}}{{f_{\max }^{nor} - f_{\max }^{ur}}}, \;\quad f_{\max }^{nor} \le f \le f_{\max }^{ur}\\ 1, \quad\quad\quad\quad\quad\quad f_{\min }^{nor} \le f \le f_{\max }^{nor}\\ \displaystyle\frac{{f - f_{\min }^{ur}}}{{f_{\min }^{nor} - f_{\min }^{ur}}}, \;\;\quad f_{\min }^{ur} \le f \le f_{\min }^{nor} \end{array} \right.$$$ (33)

 $$$\displaystyle {\mu _E}{\rm{ = }}{\mu _{E{\rm{U}} \cap f}} = \min \left\{ {{\mu _{E, U}}, {\mu _{E, f}}} \right\}$$$ (34)

 $$$\begin{split} \displaystyle &{\mu _{H, p}}\left( {{p_w}} \right) = \\ &\quad\left\{ \begin{array}{l} \displaystyle\frac{{{p_w} - p_{w, \max }^{ur}}}{{p_{w, \max }^{nor} - p_{w, \max }^{ur}}}, \quad p_{w, \max }^{nor} \le {p_w} \le p_{w, \max }^{ur}\\ 1, \quad\quad\quad\quad\quad\qquad p_{w, \min }^{nor} \le {p_w} \le p_{w, \max }^{nor}\\ \displaystyle\frac{{{p_w} - p_{w, \min }^{ur}}}{{p_{w, \min }^{nor} - p_{w, \min }^{ur}}}, \quad p_{w, \min }^{ur} \le {p_w} \le p_{w, \min }^{nor} \end{array} \right. \end{split}$$$ (35)

 $$$\begin{split} &\displaystyle {\mu _{G, p}}\left( {{p_g}} \right) = \\ &\quad\left\{ \begin{array}{ll} \displaystyle\frac{{{p_g} - p_{g, \max }^{ur}}}{{p_{g, \max }^{nor} - p_{g, \max }^{ur}}},&p_{g, \max }^{nor} \le {p_g} \le p_{g, \max }^{ur}\\ 1,&p_{g, \min }^{nor} \le {p_g} \le p_{g, \max }^{nor}\\ \displaystyle\frac{{{p_g} - p_{g, \min }^{ur}}}{{p_{g, \min }^{nor} - p_{g, \min }^{ur}}},&p_{g, \min }^{ur} \le {p_g} \le p_{g, \min }^{nor} \end{array} \right. \end{split}$$$ (36)

3.2 基于自能源的生成对抗网络

 $$$\begin{split} \min \ Ob{j^D}\left( {{\theta _D}, {\theta _G}} \right) =& -{E_{Y \in P_{\rm{data}}}}\left[{\log D\left( {X, Y} \right)} \right] - \\ &{E_{Y \in {G_\theta }}}\left[{\log \left( {1-D\left( {X, Y} \right)} \right)} \right] \end{split}$$$ (37)

3.3 训练策略

 $$$\begin{split} &\max \;Ob{j^G}({\theta _G}) = \\ &\qquad\sum\limits_{{Y_{1:T - 1}}} {{G_\theta }} ({Y_{1:T - 1}}\left| S \right.) \cdot Q_{{D_\phi }}^{{G_\theta }}({Y_{1:T - 1}}, S) \end{split}$$$ (38)

 $$$Q_{{D_\phi }}^{{G_\theta }}\left( {\left( {{Y_{1:T - 1}}, X} \right), {y_T}} \right) = D\left( {X, {Y_{1:T}}} \right) - b\left( {X, {Y_{1:T}}} \right)$$$ (39)

 $$$\left\{ {Y_{1:{T_1}}^1, \cdots, Y_{1:{T_N}}^1} \right\} = M{C^{{G_\theta }}}\left( {\left( {{Y_{1:t}}, X} \right), N} \right)$$$ (40)

 $$$\begin{split} \displaystyle &Q_{{D_\phi }}^{{G_\theta }}\left( {\left( {{Y_{1:T - 1}}, X} \right), {y_T}} \right){\rm{ = }}\\ &\qquad\left\{ \begin{array}{ll} \dfrac{1}{N}\sum\limits_{n = 1}^N {D\left( {X, Y_{1:{T_n}}^n} \right)} - b\left( {X, Y_{1:{T_n}}^n} \right){\rm{, }}&t < T\\ D\left( {X, Y_{1:t}^n} \right) - b\left( {X, Y_{1:t}^n} \right){\rm{ }}{\rm{, }}&t = T \end{array}\right. \end{split}$$$ (41)

 $$$\begin{split} {\nabla J}\left( {{\theta _G}} \right)&=\frac{1}{T} \sum\limits_{t = 1}^T {\sum\limits_{y_t} {{Q_D^{{G_\theta }}} \left( {\left( {{Y_{1:t - 1}}, X} \right), {y_t}} \right) \times } }\\ & {\nabla _\theta }\left( {{G_\theta }\left( {\left. {y_t} \right|{Y_{1:t - 1}}, X} \right)} \right)=\\ & \frac{1}{T}\sum\limits_{t = 1}^T {E_{yt \in {G_\theta }}}\Bigg[ {Q_D^{{G_\theta }}}\left( {\left( {{Y_{1:t-1}}, X} \right), {y_t}} \right) \times\\ & {\nabla _\theta }\log p\left( {\left. {y_t} \right|{Y_{1:t- 1}}, X} \right) \Bigg] \end{split}$$$ (42)
3.4 自能源参数的辨识算法

4 算例分析

4.1 自能源常规模式运行下系统参数识别

 图 6 自能源电力子系统运行状态 Figure 6 Operating state of power subsystem in WE
 图 7 自能源热力子系统运行状态 Figure 7 Operating state of heating network in WE
 图 8 自能源天然气子系统运行状态 Figure 8 Operating state of natural gas network in WE

 $$$\displaystyle\varepsilon = \frac{1}{N}\sum\limits_{i = 1}^N {\frac{{\left| {{y_i} - {x_i}} \right|}}{{{x_i}}}}$$$ (43)

 图 9 三种参数辨识方法的比较结果 Figure 9 Comparison results of three parameter identification methods

 图 10 自能源输出拟合曲线 Figure 10 Output fitting curves of we-energy

4.2 自能源不同工况下系统参数识别

 图 11 电压异常时自能源输出曲线 Figure 11 Output curves of we-energy in abnormal voltage

 图 12 液压异常时自能源输出曲线 Figure 12 Output curves of WE in abnormal fluid pressure

 图 13 气压异常时自能源输出曲线 Figure 13 Output curves of WE in abnormal gas pressure
5 结论