Fourier Series-based Kernel for Frequency Control Performance Monitoring of Hydroelectric Generating Units G. Robert12 , G. Besançon2 Abstract— An identification algorithm is proposed to estimate varying parameters and some unknown input in a MISO continuous-time system, for the purpose of monitoring frequency control performance in hydroelectric generating units. In the proposed algorithm, an integral transform with a Fourier series-based kernel is developed from irregularly sampled data and its efficiency is shown by comparison with real hydropower plant parameters and with a S avitzky-Golay filter. I. INT RODUCT ION To guarantee the frequency stability of an electric power system, power plants are requested by the power grid operator (generally the transmission system operator) to provide some ancillary services related to the system frequency control. Primary Frequency Control (PFC) and Secondary Frequency Control (SFC) refer to these services. Good explanations of the power generation operation can be found in [1]. Basically, an imbalance between production and consumption in the power network causes a frequency deviation which is compensated by control systems equipping power generation units and having PFC and SFC functions. In France, power plants which participate to ancillary services have a contractual commitment with the power grid operator to modulate the generation power with specific dynamic performance criteria (power response time, static error, damping…). These requirements become more and more severe with respect to the future European grid code [2]. The issue for the producer is to check the power plant frequency control performance, that is to assess the contribution to PFC and SFC in order to fulfill contractual commitments and possibly to detect faults in the power generation process. Few papers exist in the literature concerning the power grid frequency control performance monitoring [3] [4]. In [3], Wang uses a discrete linear time invariant (LTI) model to identify parameters related to PFC provided by a thermal generating unit. In [4], for wind turbines, again a LTI model is used for parameter estimation. None of them identify SFC parameters. To that end, a continuous -time identification approach is considered via integral transforms using Fourier-series-based kernels. This approach shows good identification results, in particular when dealing with irregularly sampled data, as those that one has to face in real situations of power plant monitoring. In section II, the considered MISO system is presented from a power generation model, and declined into a MISO linear continuous-time model with varying parameters. The problem of identification for such a continuous -time system from irregularly sampled data is addressed in section III. An application to a hydroelectric generating unit is then proposed in section IV where the unknown input is estimated as well, and section V finally concludes the paper. II. PROBLEM ST AT EMENT The nonlinear model of a power generation unit is represented hereafter by a LTV model. Varying parameters of this model are gains, time constant and damping ratio which characterize the expected dynamic behavior in compliance with grid operator requirements. A. Power Generation Model The power response P(t) of a generating unit participating to PFC and SFC is usually the sum of 3 terms: - the power response P0 (t) of the production program or a particular setpoint Pco (t), - the power response P1 (t) of the primary frequency control steered by the network frequency deviation ΔF(t) = F(t) - Fn , - and the power response P2 (t) of the secondary frequency control steered by N 2 (t ) the remote setpoint sent by the grid operator. Around an operating point, we can write deviations with Laplace formalism and transfer functions H0 , H1 , H2 : P(s) P0 (s) P1 (s) P2 (s) (1.1) P H 0 (s)Pco H1 (s) K F H 2 (s) Pr N2 (1.2) In the present paper, both PFC and SFC parameters will be estimated for control performance monitoring purposes. A hydroelectric powerplant will be considered with variant setpoints. In this context, the paper addresses the problem to estimate both an unknown input and varying parameters of a multi-input single output (MISO) system. The gain Pr represents the secondary power reserve in SFC and K is linked with the controller speed droop ratio b p : 1 2 Gérard Robert is with EDF Hydro, 73373 Le Bourget du Lac, France (gerard.robert@edf.fr). Gildas Besançon and Gérard Robert are with Univ. Grenoble Alpes, CNRS, Grenoble INP*, GIPSA-lab, 38000 Grenoble, France (gildas.besancon@gipsa-lab.grenoble-inp.fr). * Institute of Engineering K Pmax bp Fn (1.3) with Fn (Hz) the nominal network frequency and Pmax the maximu m power of the generator. The signal N2 (t) varies from -1 to 1. The output P(t) and the frequency F(t) are corrupted by measurement noise (see for instance Figure IV.1). The setpoint Pco (t) can be a piecewise constant signal (case for nuclear, thermal or reservoir hydropower plants) or a time-varying signal (run-ofthe-river power plants). The process related to power generation consists of a turbine controller (see Figure II.1), an actuator and a generating unit (turbine + power generator). The turbine controller has two inputs (R, ΔF) and two operating modes: Power control mode in which the reference signal R(t) is a power setpoint Pc and Gate opening control mode where R(t) is a gate opening setpoint. Due to nonlinear conversions in the control system, the setpoint Pc is not always available in power format. It can yet be assessed indirectly with the reference signal R(t) sent to the turbine controller. B. Equivalent LTV Model The sampling frequency of the controller is very fast compared with the acquisition of monitored variables. That’s why a continuous-time model is considered for the process model. Moreover, dynamics of the primary and secondary frequency control are approximated by a 2nd order with a natural pulsation n 1/ and a damping ratio . Regarding dynamic models for steam and hydro turbines in power system studies [6], a second order continuous -time state space representation is proposed with varying parameters which all depend on the operating point R0 : 2 x1 2 x1 x1 G1 u1 T1u1 (1.10) 2 x2 2 x2 x2 G2 u2 T2u2 (1.11) y x1 x2 (1.12) with G1 1/ bp , y P / Pmax , u1 F / Fn, u2 R / Pmax x1 P1 / Pmax and x2 P0 P2 / Pmax . From this, coefficients of the model may be poorly known since they are function of the operating point R0 . Pc (t ) Pco (t ) Pr N (t ) (1.4) Pc f ( R) (1.5) Since initial conditions are unknown, we build an inputoutput LTV model by summing the two equations (1.10) (1.11): where f is a nonlinear uncertain function 1 representing the turbine characteristic. 2 y 2 y y G1 u1 T1u1 G2 u2 T2u2 (1.13) Whatever the operating mode is, from the power system viewpoint, the following linearized model can be stated in equivalent power format around an operating point: Pc (t ) Pco (t ) Pr N2 (t ) G2R(t ) (1.6) Let us set now pco Pco /Pmax and u3 N2 /Pmax , so from (1.6) we can write: where G2 denotes a conversion gain: G2 f / R R and 0 The conventional structure of turbine controller [1] [5] allows us to write some properties for the transfer functions appearing in (1.2): H 0 ( s) H 2 ( s) (1.7) H 0 (0) H 2 (0) 1 (1.8) H1 (0) 1 (1.9) In addition, it is assumed that these 3 transfer functions have the same pair of poles. T urbine Controller Generating Unit (1.14) pc G2u2 (1.15) with where Pr and G2 depend on the operating point R0 . ΔF Δ R pc (t ) pco (t ) Pr u3 (t ) P The goal is then to estimate parameters in equation (1.13), and from them gain K will be deduced with relation (1.3). Gain Pr and the setpoint Pco(t) will be estimated in section IV.B from (1.14). The noisy measurements signals P(t), F(t) are known and the setpoints R(t) and N2 (t) are available through irregular sampling. Hence, u 1 , u 2 , u 3 and y are known. The ordinary differential equation (ODE) given by (1.13) can be rewritten in regression form: y(t ) hT (t ) (1.16) hT u1 u1 u2 u2 y y (1.17) 1 ... 6 (1.18) T Figure II.1. Power-Frequency Control of a generating unit 1 For hydropower plants, f depends also of the water head where the effect is here neglected. where the expressions of j are given in appendix. III. PARAMET RIC IDENT IFICAT ION The literature proposes many techniques to perform identification of LTV systems (see for example [7]). To estimate parameters of the LTV model, data are considered over a narrow time-shifting window [t, t+T] where parameters to identify are assumed invariant but with a time-varying unknown input Pco (t). This identification problem of such a linear continuous time-varying system is difficult for several reasons. Firstly, it is well known that calculations of the required derivatives give rise to problems of accuracy and stability [8]. Secondly, a narrow time window with an irregular sampling (i.e. a low number of samples) leads to a poor estimation quality with conventional methods [9]. Thus, a significant issue is to find an algorithm able to estimate 6 parameters within a small time window T. To generalize the problem of equation (1.13), let us consider a SISO process (extension to the MISO case will be easy) characterized by a p th order ordinary differential equation having p q 1 unknown coefficients to be identified: p q i 1 j 0 y ai y (i ) (t ) b ju ( j ) (t ) with q p, ai , b j (1.19) , u the input and y the output. The parametric identification procedure consists in transforming ODE (1.19) in a set of algebraic equations. To avoid noise amplification, many techniques exist either in frequency domain [10], in Laplace domain or in time domain [11]. For instance, we can find linear filter methods, numerical integration-based methods [10] or integral transform methods with different kernel [11] [12]. In the present paper, we will focus on an integral transform method, tailored so as to overcome problems of irregular sampling, measurement noise and eliminating the need for identifying initial conditions of the ODE. It will be compared, in the example of section IV, to an alternative filter-based approach due to Savitzky and Golay. B. Integral transform method Integral transform methods are efficient techniques to overcome noisy data. They consist in multiplying both sides of ODE (1.19) by a kernel function and integrating both sides over a time interval. Many choices are possible for the kernel: Walsh functions [11], splines functions [12], Hilbert space [16], modulating functions … A focus is made here on modulating functions since the method we will propose is similar to this one. Modulating functions-based method has been introduced in the fifties by Shinbrot [17] for aeronautics applications. Several types of modulating functions have been proposed and used, including sine functions [17], Hermit functions [18], spline-type functions [19] [12], Poisson moment functionals [20], and Hartley modulating functions [21]. The principle of modulating function technique is to transfer the continuous time derivatives signals to the derivative of a set of smooth user-chosen “modulating functions” which constitute the kernel of the integral transform applied to the ODE. The same approach is used hereafter with a different insight and a new kernel. C. Fourier series-based kernel Let us consider an integral transform I with the kernel and n : (t, n) C p where t T I [ x](n) X (n) x, (n) (t , n) x(t )dt (1.21) 0 The particularity of this transformation is that it sets up a map between a continuous time signal x (t ) and a sequence of numbers X ( n) yielding a discrete variable in the frequency domain. In order to get rid of derivatives terms appearing in (1.19), the following proposition is considered. Proposition: Let x(t ), (t , n) C p [0,T ] . If: ( j ) (0, n) ( j ) (T , n) 0 , j 0,1,..., p -1 A. Savitzky-Golay filter Developed in 1964, the Savitzky-Golay filter also known as polynomial smoothing filter is a very efficient algorithm to smooth data through a polynomial fitting and to estimate time derivatives [13]. In his book [14], Orfanidis gives a nice description of the method which can be written as a convolution equation: ys(i ) (k ) i ! M m M gi (m) y(k m) then x( p ) , (1) p x, ( p ) where x ( p ) (1.22) (1.23) d px dt p Proof: x ( p ) , (t , n) x ( p ) (t )dt and by T (1.20) with i 0,1,..., d and g i is given by a least squares algorithm [13] [14] or through the Matlab function sgolay from the Signal Processing Toolbox [15]. By definition, 0 successive integration by parts the right-hand side is equal to: T (0) x( p1) (1) x( p2) ... (1) p ( p1) x(0) 0 T The d order polynomial used in this method is calculated on a moving window with half-length M [14]. (1) p ( p ) xdt 0 Thus, (1.23) comes from boundary conditions (1.22). (1.24) Our contribution is to propose a new kernel which is a periodic function (period T/n) expanded in a finite Fourier series where coefficients will be calculated to satisfy initial and final values of condition (1.22): L (t , n) ak cos knt bk sin knt (1.25) k 0 with n n2 /T . For sake of simplicity, an even function is chosen giving a cosine partial sum: L (t , n) a0 ak cos knt (1.26) k 1 with The 2 /T . j order derivative is: L which can be determined by solving a set of with m 1,2,..., M : L ak a0 k 1 L k 2ma 0 k k 1 M 1 equations ,p (1.32) ,p 2 If the coefficient a 0 is fixed to an arbitrary value, it can be easily shown that the solution {a k} of (1.32) is proportional to a 0 . That means that the kernel is also proportional to a 0 and that this coefficient disappears when the integral transform I (1.21) is applied to equation (1.16). That’s why, we will take a0 1 . For a 1st or 2nd order dynamic system (p = 1 or 2), we find ( j ) (t , n) kn ak cos knt j j k 1 (1.27) L (t, n) 1 cos nt 2 Boundary conditions (1.22) applied to (1.26) for j 0 give: a0 ak 0 a1 a0 1 and hence: ,p (1.33) With a period T/n, this simple kernel is illustrated in Figure III.1 for n = 1 and n = 2. (1.28) k 1 And applied to (1.27) for equations: L k a j k k 1 with j 0 , we obtain a set of J cos j 0 2 (1.29) j 1,2,..., J . 0 for odd integers j , two cases must 2 be considered for p 2 : Since cos j Figure III.1: Fourier series kernel (1.35) p 2 if p 1 is odd J p 1 if p 1 is even (1.30) In addition, we set j 2m and M J / 2 in order to have the same number of equations as the number of unknown coefficients ak . So, equation (1.29) becomes: L k k 1 with 2m ak 1 0 m ,p 2 th a1 a2 a0 a1 4a2 0 (1.31) whose result is: m 1,2,..., M and L M 1 Thus to summarize, the cosine series kernel For a 3th or 4nd order dynamic system (p = 3 or 4), we have to solve: (1.34) 4 1 a1 a0 and a2 a0 . 3 3 The cosine kernel is therefore: verifying the p order relation (1.23) is characterized by coefficients ak 4 1 (t , n) 1 cos nt cos 2nt 3 3 (1.35) IV. A PPLICATION TO A HYDROELECTRIC GENERAT ING UNIT In this section, let us consider a 100 MW run-of-river hydroelectric generating unit in France as an application example for the above approach. In this example, actual parameters are: G1 = 25, G2 = 0.8, T1 = 10 s, T2 = 20 s, 12 s, 0.8 Fn = 50 Hz, Pmax = 100 MW and K = 50 MW/Hz Experimental measurement signals (see Figure IV.1) are available with a non-uniform sampling period: Ts 1, 3 s . A. Parameter Estimation Parameters are assumed invariant in a narrow window [0,T] where T (= 500 s) represents only 50 times the largest time constant (T1 ). We define the following notations: X X (n) x, (n) (1.36) X i X i (n) x, (i ) (n) (1.37) Figure IV.1: Industrial process data (1.38) In tab. I, a comparison is given between the actual parameters and parameters estimated for t 0, T by the two By applying the integral transform (1.21) with the kernel (1.33) to the equation (1.16) and using notations (1.36)(1.38), we can write : discussed methods: Savitzky-Golay (SG) and Fourier series (FS). To obtain good performance with respect to measurement noise, the SG filter is here parameterized with d = 2 and M = 41. H (n) h , (n) T n 1,..., N Y (n) H (n) (1.39) Numerical results show a better matching for the Fourier series method especially for the time constant T1 . with the regressor vector: H (n) U1 U11 U 2 U 21 Y 2 Y 1 (n) Hence we have a linear regression problem in a T-large window. Due to the number p + q + 1 of parameters to be estimated in θ, we consider a sufficiently large 2 number N of equations (1.39): Y (1) H (1) Y ( N ) H ( N ) Y H (1.41) Under sufficient excitation indeed, system (1.41) can become invertible, and an estimate of θ be obtained as: H T H H TY 1 (1.42) Finally, physical parameters (gains and time constants) are easily deduced from equations given in appendix, as well as gain K (from G1 and equation (1.3)): K G1 2 Pmax Fn TABLE I. (1.40) (1.43) Note that N should not be chosen too large either, in order to keep enough accuracy in the computation of integrals from sampled measurements. NUMERICAL RESULTS G1 G2 T1 T2 Actual 25 0.8 10 20 12 0.8 SG 24.32 0.76 3.25 18.24 9.41 0.90 FS 24.96 0.80 11.16 20.39 13.05 0.76 From relation (1.43), we find K = 2G1 that is 49.92 MW/Hz with FS method compared to the expected value 50 MW/Hz. B. Estimation of Pco(t) and Pr From estimation (1.42), the gain G2 is known and thus pc can be calculated from (1.15). The signal p co (t) is time bounded on [0, T] and we assume that pco L2 ([0, T ]) . In order to expand pco (t ) in a Fourier series (see [22]), we extend it by parity on [-T, T] then we extend it again on with the period T0 2T . Thus, with the angular frequency 0 2 / T0 / T , we have the following finite Fourier series: N0 pco (t ) ak cos k0t (1.44) k 0 For our application, we limit this expansion to 3 harmonics in (1.14) we have a linear N0 3 , and by injecting (1.44) REFERENCES [1] A. Wood and B. Wollenberg, Power generation operation and control, USA: Wiley, 1996. [2] ENT SO-E, "Load Frequency Control and Reserves Network Code," https://www.entsoe.eu/fileadmin/user_upload/_library/resources/LC FR/130628-NC_LFCR-Issue1.pdf, 2013. J. Wang, "Performance assessment of primary frequency control responses for thermal power generation units using system identification techniques," Electrical Power and Energy Systems, vol. 100, pp. 81-90, 2018. L. Castro, "A new method to assess the contribution of VSC-HVDC connectedwind farms to the primary frequency control of power networks," Electric Power Systems Research, vol. 154, pp. 48-58, 2018. N. Kishor and J. Fraile-Ardanuy, Modeling and Dynamic Behaviour of Hydropower Plants, IET , 2017. IEEE-Committee, "Dynamic models for steam and hydro turbines in power system studies,," IEEE Trans. Power Apparatus System, IEEE Committee Report, vol. 92, p. 1904–1915, 1973. J. Lataire and R. Pintelon, "Continuous-time linear time-varying system identification with a frequency-domain kernel-based estimator," IET Control Theory and Applications, vol. 11, pp. 457465, 2017. R. Isermann and M. Münchhof, Identification of Dynamic Systems: An Introduction with Applications, New York: Springer, 2010. L. Ljung, System Identification: Theory for the User, Prentice-Hall, 1998. H. Garnier and L. Wang, Identification of Continuous-time Models from Sampled Data, Springer London, 2010. N. Sinha, Identification of continuous-time system, Springer Science, 1991. Z. Liu and H. Fang, "A parameter identification method for continuous-time nonlinear systems and its realization on a Miuraorigami structure," Mechanical Systems and Signal Processing, vol. 108, p. 369–386, 2018. A. Savitzky and M. Golay, "Smoothing and differentiation of data by simplified least squares procedures," Analytical Chemistry, vol. 36, pp. 1627-1639, 1964. S. Orphanidis, Introduction to signal processing, Prentice-Hall, 1996. Mathworks, "MAT LAB, www.mathworks.com". D. Ghoshal and K. Gopalakrishnan, "Using algebraic parameter estimation kernel representation of linear systems," in IFAC WC, T oulouse, 2017. M. Shinbrot, "On the analysis of linear and nonlinear system," Trans. ASME, vol. 79, pp. 547-552, 1957. [3] regression problem computed on the interval [0, T]: pc 1 cos 0t cos 20t cos30t u3 a0 a1 a2 a3 Pr T (1.45) [4] (1.46) Least squares calculus yields which permits to reconstruct the signal Pco (t) shown in Figure IV.2. For the parameter Pr , the estimated value is 4.8 MW compared to 5 MW for the actual value. [5] [6] [7] [8] [9] [10] [11] [12] Figure IV.2: Comparison between pco actual and estimated V. CONCLUSION In this paper, a Fourier series approach has been considered to solve a MISO continuous-time LTV system identification problem and to estimate an unknown input from irregularly sampled data. Experimental data from a hydro generating unit allowed to evaluate the proposed estimator, which indeed has given a good matching between expected and estimated parameters. The proposed Fourier kernel has the advantage to be simple, easy to implement and applicable for ODEs of any order. This method is also valid for other types of power plants (wind turbine, PV, nuclear, thermal power plants). APPENDIX Physical parameters and identification parameters are linked with the following relations: 1 G1 ; 2 G1T1 ; 3 G2 ; 4 G2T2 ; 5 2 ; 6 2 (1.47) and reciprocally: G1 =1 ; G2 3 ; T1 2 / 1 ; T2 4 / 3 ; 5 ; 6 2 5 (1.48) [13] [14] [15] [16] [17] [18] K. T akaya, "The use of Hermite functions for system identification," IEEE Transactions on Automatic Control, vol. 13, pp. 446-447, 1968. [19] H. Preisig and D. Rippin, "Theory and application of the modulating function method—Review and theory of the method and theory of the spline-type modulating functions," Computers & Chemical Engineering, vol. 17, pp. 1-16, 1993. [20] D. Saha and B. Rao, "Structure and parameter identification in linear continuous lumped systems: the poisson moment functional approach," International Journal of Control, vol. 36, p. 477{491, 1982. [21] O. Cieza and J. T afur, "Frequency Domain Modulating Functions for Continuous-Time Identification of Linear and Nonlinear Systems," in 16th Latin american Control Conference, 2014. [22] R. Bracewell, T he Fourier Transform and Its Applications, Boston: McGraw-Hill, 2000.