1 High-Power High-Frequency Converter Modelling Using Dommel’s and Runge-Kutta Methods in ABC and DQ frame Wei-Xing Lin, Member IEEE and Dragan Jovcic, Senior Member, IEEE weixinglin@abdn.ac.uk, d.jovcic@abdn.ac.uk School of Engineering, University of Aberdeen, Aberdeen, UK Abstract—As the future DC grids will involve numerous converter systems, the accuracy and speed of their modeling becomes of high importance. This paper develops and compares four average models for high frequency converters like DC/DC or DC hubs: the average model in the ABC frame using Dommel’s method (Model 1), the average model in the dq frame using Dommel’s method (Model 2), the average model in the ABC frame using Runge-Kutta method (Model 3) and the average model in the dq frame using Runge-Kutta method (Model 4). Schematics and principles of the numerical algorithms for the four models are presented. Detailed switching model of a 500MW, 2kHz, ±200kV /±400kV DC/DC converter is taken as benchmark for comparing the four average models. It is found that Model2 is numerically unstable. Model 1 and Model3 need the same small solution time step as the detailed switching model to give accurate results. Model 4 is the recommended average model for high frequency (1-3kHz) converter. It gives accurate result as the detailed switching model at the benefit of 100 times faster simulation speed. Index Terms—Converter modeling, High Frequency Components, DC Power Systems, DC power transmission, DC-DC power conversion, Numerical simulation methods I. INTRODUCTION DC grids and micro grids have gained significant interest both in the industry and academic in recent years [1]-[2]. In each of these new grids, large number of power electronic converters will be used. VSC AC/DC converters will provide interface with local AC networks while AC/DC/AC converter systems provide connection to various distributed resources. As the voltage levels inside a DC grid or inside a micro grid cannot be the same, DC/DC converter will be used. In most DC/DC converter concepts there will be a two stage DC/AC/DC conversion system [3]-[4], and high frequency operation will be used in the inner AC circuit in order to reduce the size and cost. In [3] a high frequency Inductor-Capacitor-Inductor (LCL) circuit is employed while in [4] high frequency solid-state transformer is used. 50Hz/60Hz grid is 50μs which implies around 400 calculations per cycle [5]. The AC/DC converter systems operating at around 1kHz switching frequency require 2-10μs step implying 100-400 calculations per switching event. To improve the simulation speed, fast average value model (AVM) of power converters are required. The average models employed with EMTP [6],[7] are mainly focused on 50/60HZ SPWM AC/DC converters or buck/boost converters. If high frequency DC/DC converters are employed the situation becomes more complex. Here we have 1-3kHz fundamental frequency in the inner circuit and depending on the topology the converter switching frequency will be 2-5kHz. The required simulation step will be around 1μs. Simulation step for the whole DC grid must be same as for slowest converters and this becomes unacceptably slow when complex DC grids are studied. We will illustrate in this study that the average value modeling in ABC or DQ frames is not suitable for high fundamental frequency on EMTP platforms. We will examine if the default solvers used with EMTP, based on Dommel's methods [8] can be explored for high frequency component modeling. In order to provide accurate and fast models we will also employ other numerical algorithms like Runge-Kutta which are used in other simulation platforms [9][10],[11]. We further study both ABC frame and DQ frame modeling since with high frequency components these frames give much different simulation speeds and accuracy. II. DC-DC CONVERTER TESTS SYSTEM This paper takes the high power LCL DC/DC converter [3] as an example of high frequency component. Topology of the 4-phase LCL DC/DC converter is shown in Fig. 1. Rating of the DC/DC converter is ±200kV/±400kV, 500MW and its key parameters are shown in Table 2 in the Appendix. The parameters in Table 2 are designed at 550MW at a rated modulation index of 1 to provide 10% control margin. Typical simulation tool for such power electronic based grids is PSCAD/EMTDC [5] or other platform from the EMTP (Electromagnetic transients program) family. However, PSCAD/EMTDC was designed to primarily simulate 50Hz/60Hz AC grid and few AC/DC converters coupled to the grid (including HVDC systems). Typical simulation step for the electro-magnetic transients of This project is funded by European Research Council under the Ideas program in FP7; grant no 259328, 2010. Fig. 1. Circuit diagram of 4-phase high frequency LCL DC/DC converter 2 III. DOMMEL'S METHOD MODELING A. Background of Dommel’s method The Dommel’s method [8] is widely used in electromagnetic transient programs. It is based on trapezoidal rule to convert the network differential equations to algebraic equations [10]. The trapezoidal method is an implicit solver which is A-stable (stability guaranteed for any time step) [11]. However the solution is not guaranteed for the iterative process, which is involved with non-linear systems, and this indirectly affects stability. Also the accuracy of trapezoidal method is among lowest of all solvers. Fig. 2 shows the representation of inductor and capacitor in Dommel’s method. Both the inductor and capacitor are represented by an artificial resistor in parallel with controlled current source. A set of algebraic equations can then be formulated using the nodal analysis method to compute the unknown node voltages at each time t, as explained in [8]. C. AVM using Dommel’s Method in dq Frame (Model 2) AVM modeling can significantly improve simulation speed for a 50Hz/60Hz SPWM modulated VSC converter. However with ABC AVM model of the test DC/DC converter the dominant frequency of the inner circuit is 2kHz and the solution time step still needs to be small (around 1-2μs) in order to provide around 200-400 calculations per cycle. To increase the solution time step of LCL DC/DC converter we propose to represent and solve the dynamics of the high frequency inner LCL circuit in the dq rotating frame. In such case, all the electrical variables become DC variables. The dq type of modeling is used frequently in order to develop converter controllers. On the downside dq modeling will not properly represent unbalanced AC system events. In a dq frame, the dynamics of inductor and capacitor are: dIid 1 = (− Ri I id + ωo Li Iiq + Vid − Vcd ) dt Li dIiq Fig. 2. Representation of inductor and capacitor in Dommel’s method B. AVM using Dommel’s Method in ABC Frame (Model 1) The common way of AVM modeling for VSC converters is to neglect the switching dynamics and use controlled AC voltage source at fundamental frequency [6], and this method will be used for modeling complex DC grids [7]. Fig. 3 shows the DC/DC average model in the ABC frame. The two VSC converter bridges in Fig. 1 are replaced by two 4-phase controlled AC voltage sources (v1A-v1D and v2A-v2D). The controlled AC voltage sources are given as: v1 A = m1AV1dc , v1B = m1BV1dc , v1C = m1CV1dc , v1D = m1DV1dc (1) Where m1A-m1D are modulation indices in the ABC frame. The inner LCL circuit of Fig. 3 will be solved using graphical interface in PSCAD /EMTDC, since PSCAD /EMTDC uses only the Dommel’s method [6]. = dt dVcd 1 = ( I1d + I 2d + ωo CVcq ) dt C dVcq 1 = ( I1q + I1q − ωo CVcd ) dt C (3) Using constant amplitude dq transformation, active power of a Mp-phase VSC converter at the AC side is Paci = Mp 2 (Vid I id + Viq I iq ) = Mp 2 kmVidc ( M id I id + M iq I iq ) (4) = Pdci = 2Vidc Iidc Therefore, the AC/DC interface equation is Iidc = Mp 4 km (M id I id + M iq I iq ) (5) where km is a coefficient representing the modulation method. km approximates to 4/π for the modulation method adopted in the simulations in this paper and used in [3]. Fig. 4 shows the schematic of DC/DC average model in the dq frame, and Dommel’s solving method is used since model is implemented in PSCAD/EMTDC. An artificial d-axis circuit and q-axis circuit are built in PSCAD/EMTDC to represent the state-space equations of (2)-(3). Relationships between Vid, Viq and the modulation indexes Mid and Miq are: Vid = M id Vidc Viq = M iqVidc Fig. 3. Average Model using Dommel’s method in ABC frame (Model1) (2) 1 (− Ri I iq − ωo Li Iid + Viq − Vcq ) Li (6) The power balance between AC and DC systems is represented by controlled current sources, using (5) as shown in Fig. 4. 3 Fig. 4. Average model using Dommel’s method in dq frame(Model 2) IV. RUNGE KUTTA SIMULATION A. Background of Fourh Order Runge-Kutta Method The Dommel’s method is based on trapezoidal solver, where the order of its global truncation error is proportional to square of simulation step (O(h2)). The Runge-Kutta method is a high precision method for solving differential equations. Order of its global truncation error is proportional to fourth order of simulation step (O(h4)), and therefore accuracy is better. On the downside Runge-Kutta approach belongs to explicit methods which are not A-stable and therefore instability will occur for a sufficiently large time step [11]. If time step is sensibly selected, the Runge-Kutta method will give both: good stability and accuracy and it is widely employed for simulation of engineering systems [9]-[11]. A general nonlinear state space model dx/dt=f(x,u) can be solved using the Runge-Kutta’s method in the following steps: 1) calculate k1 = Δtf ( x (t ), u(t )) 2) calculate x1 = x (t ) + 0.5k1 3) calculate k2 = Δtf ( x1 , u(t )) 4) calculate x2 = x (t ) + 0.5k2 5) calculate k3 = Δtf ( x2 , u(t )) 6) calculate x3 = x (t ) + k3 7) calculate k4 = Δtf ( x3 , u(t )) 8) x (t + Δt) = x (t) +1/ 6(k1 + 2k2 + 2k3 + k4 ) Fig. 5 shows the schematic of DC/DC average model in the ABC frame using Runge-Kutta solving method. As the PSCAD/ETMDC has only one solving method, a user defined solving algorithm needs to be developed in order to implement the Runge-Kutta method. The user model takes L, R and C as the input parameters and the modulation indexes m1A-m1D and m2A-m2D, E1 and E2 as the input variables. For every simulation step, the model will calculate the state variables following the procedures of IV.A. The outputs of the user model are the inductor currents and capacitor voltages in the ABC frame and the DC currents interfacing the DC system. These variables are then used as inputs for the controllers in the next simulation step. Fig. 5. AVM of Runge Kutta method in ABC frame (Model 3) C. AVM using Runge-Kutta Method in dq Frame (model 4) Fig. 6 shows the schematic of DC/DC average model in the dq frame using Runge-Kutta solving method. It is quite similar to Fig. 5 except that the state-space equations of (2)-(3) instead of (7) will be used. Assuming symmetrical and balanced system, a number of phases can be converted to DQ frame with only two variables per dynamic element. The advantage is that there are 12 states for a 4-phase DC/DC converter model in the ABC frame while only 6 states in the dq frame. B. AVM using Runge-Kutta Method in ABC Frame(Model 3) While first order differential equations are employed in the Dommel’s method, the Runge-Kutta’s method is based on state-space models. The state-space model of one phase of the inner LCL circuit is: Fig. 6. AVM of Runge Kutta method in ABC frame(Model 4) 0 ⎡i1A ⎤ ⎡− R1 / L1 d ⎢ ⎥ ⎢ i = − R 0 2A 2 / L2 dt ⎢ ⎥ ⎢ ⎢⎣vcA ⎥⎦ ⎣⎢ 1/ C 1/ C −1/ L1 ⎤ ⎡i1 A ⎤ ⎡1/ L1 0 ⎤ ⎡ m E / 2 ⎤ (7) −1/ L2 ⎥⎥ ⎢⎢i2 A ⎥⎥ + ⎢⎢ 0 1/ L2 ⎥⎥ ⎢ 1 A 1 ⎥ m E /2 0 ⎦⎥ ⎢⎣vcA ⎥⎦ ⎣⎢ 0 0 ⎦⎥ ⎣ 2 A 2 ⎦ Similar as the derivation of (5), the DC current is calculated by: i1dc = 2 π (M1Ai1 A + M1B i1B + M1C i1C + M1D i1D ) (8) V. MODEL TESTING A. Numericals stability The detailed switching model of the 2kHz LCL DC/DC converter with solution time step of 1us is used as benchmark for comparing the performances of the four average models. It is found that Model2 is numerically unstable. Large artificial resistors in series with the inductors (R1 and R2 in Fig 4) are required to make Model2 numerically stable. 4 Fig. 7 compares the output of Model2 with the outputs from detailed switching model. Power order is stepped from 0 to 1pu at 1.0s, which is then reversed to -1.0pu at 2.0s, and a permanent DC fault is applied at 3.0s. We can see from Fig. 7 that Model2 requires artificial resistor to make it stable. A 30Ω resistor will make Model2 numerically stable but this resistance is unrealistic, and there is significant steady state error between Model2 and the detailed switching model. Although this modeling approach is adequate with 50Hz AC/DC converters, it is found to be inadequate for any higher frequency converters. 600 Pdc1(MW) 400 Model2-30Ω 200 0 Model2-20Ω -200 -400 Detail -600 0.5 1 1.5 2 Time(s) 2.5 3 3.5 Fig. 7. Numerical instability of Model2 idc2(kA) B. Accuracy Fig. 8 compares the accuracy of the Model1, Model3 and Model4 with the detailed switching model. The applied system disturbance is same as in Fig. 7 but DC current is monitored. 0.8 0.6 0.4 0.2 0 -0.2 -0.4 -0.6 -0.8 Detail Doml-ABC-1us Doml-ABC-10us 0.5 1 1.5 2 2.5 Time(s) 3 3.5 4 idc2(kA) (a) Dommel’s Method in ABC frame(Model1) 0.8 0.6 0.4 0.2 0 -0.2 -0.4 -0.6 -0.8 Detail Rung-ABC-1us Rung-ABC-10us 0.5 1 1.5 2 2.5 Time(s) 3 3.5 4 idc2(kA) (c) Runge-Kutta Method in ABC frame(Model3) 0.8 0.6 0.4 0.2 0 -0.2 -0.4 -0.6 -0.8 1 1.5 2 2.5 Time(s) 3 3.5 Fig. 8(b) shows that behaviour of Model3 is almost the same as the behaviour of Model1. Model3 is able to generate similar result as the detailed switching model only with solution step of 1μs. Fig. 8(c) shows that the output from Model4 is well matching the detailed switching model. With this model large solving time step can be adopted, and even 50μs gives excellent accuracy. C. Model final comparisons Table 1 summarizes the overall comparisons of the 4 average models. Solution time step for Model1-Model3 is 1μs while it is 50μs for Model4. The maximum allowable simulation step for Model1 and Model3 to give similar result as the detailed switching model is 1μs. The maximum time step is 50μs for Model4. Table 1 shows that Model4 can give accurate results at the benefits of increasing the simulation speed by nearly 100 times. Table 1 Overall comparison of the 4 average models for 2kHz DC/DC Converter Simulation Accuracy is Solution Numerical Models time for 4s good for Method stability real time time step Detailed Model stable 216.3s ≤1μ s Model 1 (ABC, Domel 201.3 stable ≤1μ s 1us) Method Model 2 (DQ, \ \ unstable 1us) Model 3 65.8 stable Runge≤1μ s (ABC,1us) Kutta Model 4 Method stable 2.1 ≤50μ s (DQ,50us) D. Influence of operating frequency The above four models are further tested on a reduced frequency (200Hz) LCL DC/DC converter. Voltage and power ratings are the same with the 2kHz LCL DC/DC converter. Therefore, the AC inductors and capacitors of the inner LCL circuit will be increased by 10 times. Model2 is still numerically unstable with 200Hz converter and will not be further discussed. Power order at terminal1 is stepped from 0pu to 1.0pu at 1.0s and again stepped down from 1.0pu to -1.0pu at 3.0s. A permanent DC fault is applied at terminal 2 at 3.5 s. Fig. 9 shows outputs of the 3 models using different simulation steps. As the operating frequency is reduced from 2kHz to 200Hz, the simulation step for the detailed model is increased from 1us to 10μ s. Detail Runge-dq-10us Runge-dq-50us 0.5 Fig. 8(a) shows that Model1 is able to generate similar results as the detailed switching model, but only with the small solution time step which is similar as the step in the detailed switching model. When the solution step is increased up to 10μs, accuracy of Model1 deteriorates significantly. 4 (c) Runge-Kutta Method in dq frame(Model4) Fig. 8. Accuracy of Model1, Model3 and Model4 for 2kHz DC/DC Converter Conclusions for the 3 models are the same as in the 2kHz case. Model 1 and Model3 can generate similar results as the detailed switching model in both normal operation condition and during DC faults if the solution step is the same as the 5 detailed model. Model4 is able to generate almost the same result as the detailed switching model at significantly increased time step and improved simulation speed. Fig. 9 (d) gives a very interesting result. Model4 is still able to generate almost the same result as the detailed switching model even when the solution time step is as high as 500μs, which testifies robustness of the proposed modelling approach. We can envisage the solution step of a DC grid could be larger than 50μ s if all the related AC systems are modelled in the dq frame. 1000 E1,E2(kV) 800 600 400 E1 E2 200 0 -200 2.9 3 3.1 3.2 3.3 Time(s) 3.4 3.5 3.6 (a) DC voltages 1 Detail-10us Doml-ABC-10us Doml-ABC-50us idc2(kA) 0.5 0 VI. CONCLUSION Four average models for high frequency converters are developed and tested. They are the Dommel’s method in the ABC frame (Model1), the Dommel’s method in dq frame (Model2), the Runge-Kutta Method in the ABC frame (Model3) and the Runge-Kutta method in the dq frame (Model4). It is concluded that Model2 is numerically unstable, Model1, Model3 and Model4 can generate similar results as detailed switching model in both normal operation and during large transients such as DC faults. However simulation speed of Model1 and Model3 is similar as the detailed switching model. Simulation speed of Model4 is around 100 times faster than the detailed switching model. Model4 is recommended for modeling DC grids and micro grids in which high frequency components such as DC/DC converter or DC hub will be included. Model4 is accurate enough to replace detailed switching model in electromagnetic study of DC grids, while the simulation speed is hundreds of times faster than the detailed switching model. The DQ axis Runge-Kuta modeling is expected to become more complex for large DC grids, which will become topic for our future research. -0.5 VII. APPENDIX -1 Table 2 Parameter of a 2kHz LCL DC-DC converter -1.5 2.9 3 3.1 3.2 3.3 Time(s) 3.4 3.5 3.6 V1dc(kV) V2dc(kV) L1(H) L2(H) C(uF) Prate(MW) f(Hz) ±200 ±320 0.0409 0.0498 0.1667 500 2000 (b) Dommel’s Method in ABC frame(Model1) 1 0 idc2(kA) Table 3 Parameter of a 200Hz LCL DC-DC converter Detail-10us Rung-ABC-10us Rung-ABC-50us 0.5 -0.5 L1(H) L2(H) C(uF) Prate(MW) f(Hz) ±200 ±320 0.409 0.498 1.667 500 2000 D Jovcic, K.Linden, D. Van Hartem, J.P. Taisne “Feasibility of DC transmission Networks” ISGT Europe, Panel session proceedings, Manchester, December 2011. [2] H. Tao, A. Kotsopoulos, J. L. Duarte and M. A. M. Hendrix, “Family of multiport bidirectional DC-DC converters,” IEE Proc.-Electr. Power Appl., vol. 153, no. 3, pp. 451-458, May. 2006. [3] D Jovcic, and L Zhang, “LCL DC/DC converter for DC grids” IEEE Trans. Power Del., vol 28, no. 4, 2013, pp 2071-2079. [4] S. Falcones, R. Ayyanar and X. Mao, “A DC-DC multiport converter based solid state transformer integrating distributed generation and storage,” IEEE Trans. Pow. Elec. vol. 28, no. 5, pp. 2192-02, May 2013. [5] Manitoba-HVDC Research Centre, User’s guide on PSCAD, 2005. [6] S. Chiniforoosh, J. Jatskevich, A. Yazdani, V. Sood, V. Dinavahi, J. A. Martinez and A. Ramirez, ‘‘Definitions and applications of dynamic average models for analysis of power systems,’’ IEEE Trans. Power Del., vol 25, no. 4, pp. 2655-2669, Oct. 2010. [7] T K Vrana, Y Yang, D Jovcic, S Dennetière, J Jardini, H Saad, ‘The CIGRE B4 DC Grid Test System’’, ELECTRA vol. 270, Oct. 2013, pp 10-19. [8] H. W. Dommel, ‘‘Digital computer solution of electromagnetic transients in single- and multiphase networks,’’ IEEE Trans. Power App. Syst., vol. PAS-88, no.4, pp. 388-399, Apr. 1969. [9] Dessaint, L.-A.; Al-Haddad, K. ; Le-Huy, H. ; Sybille, G. " A power system simulation tool based on Simulink" IEEE Trans. Ind. Electron., vol. 46, no. 6, pp. 1252-54, Dec 1999. [10] L. V. D. Sluis, “Transients in power systems,” John Wiley&Sons Ltd, pp. 137-139, 2001. [11] T. Hartley, G. Bealey, S. Chicatelli “Digital Simulation of Dynamic Systems : A Control Theory Approach” PTR Prentice Hall, 1994 [1] -1.5 2.9 3 3.1 3.2 3.3 Time(s) 3.4 3.5 3.6 (c) Runge-Kutta Method in ABC frame(Model3) 1 Detail-10us 0.5 idc2(kA) V2dc(kV) REFERENCES -1 Rung-dq-50us 0 Rung-dq-500us -0.5 -1 -1.5 2.9 3 3.1 3.2 3.3 Time(s) 3.4 3.5 3.6 (d) Runge-Kutta Method in dq frame(Model4) 0 idc2(kA) V1dc(kV) -0.5 -1 -1.5 3.49 Detail-10us 3.51 3.53 Rung-dq-50us 3.55 3.57 3.59 3.61 Time(s) 3.63 3.65 (e) High precision version of Fig. 9(d) from 3.49s to 3.65s Fig. 9. Accuracy of Model1, Model3 and Model4 for 0.2kHz Converter