Output Power Control Using Artificial Neural Network for Switched Reluctance Generator

We propose an output power control of a variable-speed switched reluctance generator (SRG) by implementing an artificial neural network (ANN) in the control loop. In the high-speed operation with single pulse mode, the phase current waveform, and subsequently, the output power, depend on the conduction angles. The conduction angles, i.e., the turn-on and turn-off angles, can be determined by the proposed method using an ANN. A dynamic model of an SRG with eight stator poles and six rotor poles is used for simulation to obtain the output power profiles, which subsequently become the ANN training data. The inputs of the ANN are the reference value of the output power and the rotor speeds, while the outputs of the ANN are the turn-off and turn-on angles. The control algorithm is implemented by integrating the trained data into the dynamic model using MATLAB. The experimental setup of the SRG is implemented using a digital signal processor (DSP) to control the two-switches-per-phase drive system, which includes highly accurate phase current and dc-link voltage sensor circuits. The trained biases and weights of the ANN are also coded in the DSP. To validate the proposed method, comparisons are made between simulation and experimental results.


Introduction
In power generation, a switched reluctance generator (SRG) serves as a good alternative to a DC generator. SRGs are useful in a variety of applications, including the automotive industry, (1)(2)(3)(4) wind turbines, (5) and aerospace power systems. (6) The main characteristics of a typical SRG are the simplicity of the structure and low-cost power generation with high efficiency and torque over a wide range of speeds. Nonetheless, the doubly salient pole structure and core saturation effect of SRGs contribute to highly nonlinear magnetic characteristics and more torque ripples. The output power profiles of SRGs are also subjected to various factors, including the dc-link voltage, dc-load current, phase voltage, phase inductance, rotor position, and speed. Thus, the design of an optimal output power control for an SRG that accounts for all the influencing factors presents developers with challenges.
Various SRG output power control methods have been proposed. The energy conversion process and the structures of the speed and power controllers of an SRG were proposed in Ref. 7. In Ref. 8, the optimized performance in the current-controlled mode was investigated. In Ref. 9, the output power profiles of an SRG in various applications were determined on the basis of the control variables and design, while in Ref. 10, a novel control system for variable-speed wind energy applications was discussed. A modeling method for the maximum power control of an SRG in wind turbines was presented in Ref. 11. In Refs. 12 and 13, a control strategy was proposed to enhance the output power by introducing a freewheeling mode to boost the phase current when switches were turned off. In Ref. 14, the optimal efficiency control of an SRG was determined from the optimal turn-on and turn-off angles in the single pulse mode, ignoring the nonlinear magnetization curves. In Ref. 15, a closed-loop control for high-speed SRGs was presented, whereby the optimal turn-off and turn-on angles were respectively obtained from an analytical curve fit and closed-loop power control. In Ref. 16, a control strategy by which the turn-on and turn-off angles were varied to achieve the optimal power and maximum efficiency through reductions in the dc-link ripple current was proposed. In Ref, 17, a closed-loop power control to achieve higher efficiency in SRGs, in which the optimal turn-on and turn-off angles were selected on the basis of given reference powers and rotor speeds, was investigated. Furthermore, there have been an increasing number of studies on applying an artificial neural network (ANN) in SR machine controls, such as in sensorless control, (18)(19)(20) optimized torque control, (21) speed control, (22) and wind turbine control. (23) However, there have been no studies on an output power control that incorporates an ANN. We thus propose an output power control for an 8/6 SRG using an ANN to estimate the conduction angles according to the speed and output demand. In the proposed method, the output power profiles of the SRG at several operation speeds served as the ANN training data sets.
The organization of this paper is as follows. Section 1 is the introduction. Section 2 explains the details of the output power equations. Section 3 shows the details of the ANN-based output power control scheme. Next, Sect. 4 deals with comparisons between simulation and experimental results to verify the proposed method. Finally, Sect. 5 concludes the paper.

Output Power Equation of SRG
The SRG waveforms when running in single-pulse-mode operation are illustrated in Fig. 1. The electric power P ph in a given phase is the sum of the powers in the generation and excitation stages. The generation and excitation powers of an SRG in the single pulse mode are derived as where P ph is the power of a given phase, P exc is the excitation power, P gen is the generated power, i ph are phase currents (i.e., i a , i b , i c , i d ), m is the number of phases, θ rr is the rotor pole arc, and T is the duration of one phase operation. The input power P in of the SRG drive is the product of the average torque T m and the rotor speed ω r . (24) The average output power P out can be calculated by multiplying the dc-link voltage v dc by the dc-load current i L . The input power and output power are expressed as follows:

Output Power Control Using ANN
In single-pulse-mode operation, the output power of an SRG depends on the phase current waveform, which in turn depends on the turn-off angle θ off and turn-on angle θ on and the rotor speed ω r . It is thus possible to control the variable-speed output power by varying the turn-off and turn-on angles. In this output power control scheme design into which an ANN was incorporated, the SRG output power profiles at various operation speeds are very important. Therefore, in this section, we propose how to create output power profiles.  based on the magnetizing data from finite element analysis (FEA) was simulated to determine the output power profiles at various speeds and conduction angles. (25) In Fig. 2, the conduction angle commands (θ on * , θ off * ) were adjusted to generate the optimal electric output power. The output power was calculated by multiplying the rms value of the dc-link voltage v dc by that of the dc-load current i L . In the experiments, the speed was varied between 5000 and 7000 rpm, and iterative searches for the output power profiles were performed at speeds of 5000, 6000, and 7000 rpm. Figure 3 illustrates the simulated output power profiles at the three speeds. In Fig. 3(a), in which the SRG operated at a speed of 5000 rpm, the turn-on and turn-off angles were 49 and 12°, respectively, to obtain an output power of 1800 W (P out = 1800 W). In   , when the SRG operated at a speed of 6000 rpm, the turn-on and turn-off angles were adjusted to 52 and 10°, respectively, to obtain an output power of 1800 W (P out = 1800 W). In Fig.  3(c), where the speed was 7000 rpm, the turn-on and turn-off angles were respectively 50 and 9° to maintain P out = 1800 W. To automate the tuning of the conduction angles in response to a variable speed, we incorporated an ANN into the dynamic model for automatic tuning of the conduction angles to maintain the output power at the load.

Multilayer neural network
The proposed ANN structure, shown in Fig. 4, was utilized to define the optimal turn-off angle θ off ANN and turn-on angle θ on ANN corresponding to the output power P out and the variable speed ω r of an SRG. (26) Such a network comprised three layers, i.e., the input layer (R), the hidden layer (S), and the output layer (T). The inputs of the ANN consist of the output power P out and speed ω r . The output power was calculated by multiplying the dc-link voltage v dc by the load current i L . The rotor speed ω r was calculated by counting the pulse of the incremental encoder. The input layer's activation function was a tangent sigmoid transfer function. In the hidden layer, the relationship between the inputs and outputs of the input layer can be expressed as The network outputs comprise the turn-off angle θ off ANN and turn-on angle θ on ANN . A linear transfer function is chosen as the activation function in the output layers. The relationship between inputs and outputs in the output layers is shown by the following equations:

Training data sets of the ANN
In this research, the output powers at various turn-on and turn-off angles were collected from the simulated dynamic model, as depicted in Fig. 2. They were used as the training data sets. The simulations in single pulse mode were carried out at the selected speeds of 5000, 6000, and 7000 rpm. The training data sets contained the relationships between the output power and the conduction angles at these speeds. The conduction angles were adjusted within the optimal area to optimally generate the electric output power. The parameters of the 8/6 SRG are given in Table 1; the turn-on angle θ on * can be adjusted from 45 to 60° before the aligned position. Also, the turn-off angle θ off * can also be adjusted between 1 and 15° after the aligned position. As shown in Fig. 5(a), the training data set at each selected speed consisted of 225 operation points, giving rise to 675 points for the three selected speeds. Figure 5(b) illustrates the targeted turn-off angle θ off (Target) and turn-on angle θ on (Target) of the neural network.

Feed-forward algorithm
The output power profiles were input to train a multilayer feed-forward algorithm using the MATLAB TM Neural Network Toolbox. (27) For each number of neurons, the network was trained multiple times, and the optimal number of neurons in the hidden layer (S) was 10 neurons. The matrix size of the input weight IW 1,i was thus [10 × 2], while the bias b 1 matrix size was [10 × 1]. The output layer (T) consisted of two layers: the first layer estimated the turn-on angle θ on ANN and the second layer estimated the turn-off angle θ off ANN . The weight matrices of the output layers were LW 2,1 and LW 3,1 , which were both [1 × 10] matrices. Additionally, the output layers contained b 3 and b 2 biases for the turn-off and turn-on nodes, respectively. The relevance between the inputs and outputs of the ANN were obtained using Eqs. 6-11.

Backpropagation learning algorithm
To train the ANN, the output power profiles (Fig. 3) were reformatted into training data sets [ Fig. 5(a)]. The targeted conduction angles, θ on (Target) and θ off (Target) , were then selected [ Fig. 5(b)].
The selected target turn-on angles were between 45 and 60° before the aligned position (θ a = 60, 0°), whereas the chosen target turn-off angles were between 1 and 15° beyond the aligned position. The Levenberg-Marquardt backpropagation learning algorithm was next utilized to adjust the input weight IW 1,i and the layer weights LW 2,1 and LW 3,1 of the ANN.

ANN-incorporated output power control scheme
The optimal biases and weights of the ANN were acquired at the end of the training. The results were then integrated into the dynamic model of the 8/6 SRG. Figure 7 depicts the model with the ANN-incorporated output power control. The model was developed in MATLAB/ Simulink. Hysteresis current control was used to regulate the phase current. At either a low or medium speed, the output power controller of the SRG was not in use. However, when operating in the single pulse mode, the controller was activated. The phase angles were used to control the converter and were compared with the turn-off and turn-on angle commands determined by the ANN. The speed ω r was converted into a rotor angle θ m by integration before separation into four phase angles, θ a , θ b , θ c , and θ d .
Initially, there was no output power at the load (P out = 0 W), while the output power error P out (Error) was identical to the command value P out * . The initial conduction angles were determined on the basis of the command values of speed ω r and power P out * . If the speed and output power commands were 6000 rpm and 1800 W, the ANN-calculated turn-on angle θ on ANN and turn-off angle θ off ANN would be 52.14 and 9.77°, respectively. While in operation, all of the phase currents (i a , i b , i c , i d ) flowed in the dc bus as a dc-link current i dc . The dc-link current was separated into currents of i Ce and i L to charge the capacitor and to supply the dc-load, respectively. The dc-link voltage v dc and dc-load current i L were fed back to compute the average output power P out . The hysteresis power control module was used to regulate the output power. Then, the average output power and speed became the inputs of the ANN. The outputs of the ANN were the turn-off and turn-on angles, which were optimized to the speed and output power commands. In the generating mode, the electromagnetic torques of all phases (i.e., T a , T b , T c , T d ) are negative, while the air gap torque T e is the sum of T a , T b , T c , and T d . In this research, the moment of inertia of the rotor and the stay loss were excluded.

Simulation and Experimental Results
Simulations were performed in MATLAB, and experiments were implemented using a TMS320F2812 digital signal processor (DSP) to control the system. To evaluate the proposed output power control scheme, the simulation and experimental results were compared. The SRG parameters used in the experiment are presented in Table 1. Figure 8(a) depicts a schematic of the SRG experimental system, which consisted of an induction motor (ABB) as the SRG prime mover, whose speed was in turn regulated by a threephase inverter (NOVEM). IGBT switches (PIOLAX) were assembled into a four-phase asymmetrical converter, and a three-phase diode bride rectifier was used as the dc source to excite the phase windings. In addition, an SS-200 torque transducer and a TS-2700 torque converter (ONO-SOKKI) were applied to determine the average torque T m of the SRG. The SS-200 transducer was set between two rotor shafts of the SRG and the induction motor. The rotor position and rotation speed of the SRG were determined with an R137 incremental encoder (Gurley Precision Instruments). The system also included a dc-load R L as the resistive load, EXB850 gate drive circuits (Fuji Electric), and L08P150 current sensors (Tamura). Figure 9(a) shows the 8/6 SRG used in the experiment with a torque transducer, the induction motor as a prime mover, and the incremental encoder. Figure 9(b) shows the two-switches-per-phase converter and its controller. C-language was used for DSP programming. The structure of the ANN was also written in the DSP. Figure 10  As shown in Fig. 8(b), the switching frequency ( f sw ) of the IGBTs was set at 10 kHz, resulting in the interrupt service routine (ISR) taking place every 100 μs. Once the ISR occurred, a 12-bit ADC module sampled the input variables, i.e., v dc , i L , i a , i b , i c , i d , and v a, and the entire sampling process took approximately 0.4 μs. The encoder pulses were then counted by a QEP module, and the counting and calculation of the rotor speed took about 0.6 μs. Then, the encoder pulses were converted into rotor positions, which required approximately 0.22 μs. The phase angles were next calculated, which required around 0.5 μs. The output power was computed by multiplying the dc-link voltage v dc and dc-load current i L , which took about 1.25 μs. The calculation of the conduction angles by the ANN required around 53.45 μs, and the ANN-estimated conduction angles were then applied to the gate drive module. Finally, the calculated values, i.e., the rotor position θ m , phase angles (θ a , θ b , θ c , θ d ), command angle com A * , and ANN-estimated conduction angle com A ANN , were sent to the ADC ports of the DSP for measurement. As shown in Fig. 8(b), all the computational steps consumed about 59.32 μs with an idle time of 48.68 μs. Figure 9(a) shows a test bench mounted with the 8/6 SRG, torque transducer, induction motor, and incremental encoder, while Fig. 9(b) is a photograph of the control and power circuit of the SRG drive system.

Phase current sensor circuits
To measure the phase currents (i a , i b , i c , and i d ) of the proposed SRG and the dc bus current (i L ), current sensor circuits were developed. Tamura L08P150D15 Hall effect current sensors were used in the current sensor circuits. Figure 11 depicts the circuit diagram of the current sensor circuits. As shown in Fig. 9(b), the current sensor circuits are installed in the control and power circuits of the SRG drive system. The phase currents (i a , i b , i c , and i d ) and dc bus current (i L ) waveforms will be used to verify the accuracy of the phase current sensor circuits.

DC bus voltage sensor circuit
A dc bus voltage sensor circuit was developed to measure the dc bus voltage (v dc ) of the proposed SRG. An ISO124 voltage sensor was applied in the dc bus voltage sensor circuit. Figure 12 depicts the circuit diagram of the dc bus voltage sensor circuit. As shown in Fig. 9(b), the dc bus voltage sensor circuit was installed into the control and power circuits of the SRG drive system.

Incremental encoder circuit
To measure the rotor position (θ m ) of the proposed SRG, the interface circuit between the incremental encoder and the DSP board was developed. An R137 incremental encoder was used  to obtain the rotor position and rotation speed of the SRG. Figure 13 depicts the interface circuit of the incremental encoder and DSP for calculation of the rotor position. As shown in Fig. 9(a), the current sensor circuits were equipped in the control and power circuits.

Off-line training of ANN-based 8/6 SRG drive
As shown in Fig. 10(a), an off-line training algorithm of the SRG drive system was introduced to train the weights and biases in the DSP. In the training, a turn-off angle θ off * of 10° and a turnon angle θ on * of 50°, were initially selected as the command values. The phase current command i a * was set at a high level of 100 A to make the drive system operate in the single pulse mode.
Meanwhile, the speed command ω r was controlled by the prime mover. The dc-link voltage v dc and load current i L were fed back to determine the output power. The calculated output power P out and speed ω r were the inputs of the ANN. The calculated turn-on angle θ on ANN and turn-off angle θ off ANN were compared with the command values (i.e., 50 and 10°). As shown in Fig. 14, the conduction angle errors (θ on (Error) , θ off (Error) ) were indicators of the accuracy performance of the proposed ANN. To train the weights and biases in the DSP, the speed of the drive system was varied between 5000 and 7000 rpm with an increment of 200 rpm. The turn-off angle θ off * and turn-on angle θ on commands were set at 10 and 50°, respectively, to generate the output power optimally. The load current and dc-link voltage were then fed back to compute the output power, which became an input of the ANN. The ANN then calculated the optimal conduction angles (θ on ANN , θ off ANN ) as its outputs from the selected speed and output power.

Constant-speed performance
To evaluate the proposed system under the steady-state condition, the rotor speed was fixed at 6000 rpm and then at 7000 rpm. The output power command was also fixed at 1800 W at both operation speeds. At the speed of 6000 rpm, the optimal turn-on angle θ on ANN and turn-off angle θ off ANN computed by the ANN were 52.14 and 9.77°, respectively. Figure 15(a) depicts the simulation results of the phase inductances (L a , L b , L c , L d ) and commutation angle of phase-A (com A ). The ANN-calculated turn-on angle was computed just before the aligned position at θ a = 52.14°, whereas the turn-off angle was computed by the ANN after the aligned position at θ a = 9.77°. Figure 15(b) illustrates the simulation results of the dc-link voltage v dc , dc-load current i L , and the four-phase current (i a , i b , i c , i d ) waveforms at 6000 rpm. The average simulated dc-link voltage and dc-load current were 47.12 V and 38.77 A, respectively, and the average output power at the load was 1827 W with an error of approximately 1.55%. Figure 16(a) depicts the experimental results at the same speed. The dc-load current (i L ) and average measured dc-link voltage (v dc ) were 39.21 A and 47.45 V, respectively. The average output power at the load was 1860 W with an output power error of 3.33%. The simulated and measured waveforms exhibited similar trends. At the speed of 7000 rpm, the optimal turn-on angle θ on ANN and turn-off angle θ off ANN computed by the ANN were 53.51 and 9.35°, respectively. Figure 17(a) depicts the simulation results of the phase inductances (L a , L b , L c , L d ) and commutation angle of phase-A (com A ). The ANN-calculated turn-on angle was before the aligned position (θ a = 53.51°), while the turn-off angle computed by the ANN was beyond the aligned position (θ a = 9.35°). Figure 17

Performance at a variable speed
During the variable-speed operation, the conduction angles were varied according to the speed while the output power remained constant. Initially, the SRG was driven by the prime mover at 7000 rpm. The command of the output power was fixed at 1800 W. At this speed, the ANN-calculated conduction angles (θ on ANN , θ off ANN ) were 53.51 and 9.35°.
The SRG was operated under a steady state (i.e., 7000 rpm) for 1.38 s before the speed of the prime mover was reduced to 6000 rpm while the output power remained constant at 1800 W. At 6000 rpm, the ANN recalculated the conduction angles (θ on ANN , θ off ANN ) and the new conduction angles were 52.14 and 9.77°, respectively. The SRG was again operated under a steady state (i.e., 6000 rpm) for 2.21 s before increasing the speed of the prime mover to 7000 rpm, while the output power command remained unchanged at 1800 W. At 7000 rpm, the conduction angles returned to 53.51 and 9.35°. The simulated waveforms of the phase current, phase voltage, dc-link voltage, and dc-load current under the variable-speed condition are illustrated in Fig. 19(a). Figure 19(b) shows the average output power at the load.   However, a dc-link voltage v dc ripple and load-current ripple i L ripple are noticeable in the experiments when the SRG was operated under a steady state and constant speed, a phenomenon that induced disparities between the simulated and measured power outputs. Table 2 presents the errors of the dc-link voltage, dc-load current, and average output power at the speeds of 5000, 6000, and 7000 rpm. Figure 20 illustrates the relationships between the optimal conduction angles and the average output power determined by the proposed output power control at 5000, 6000, and 7000 rpm. It can also be observed that the range of the conduction angles is wide for high output power demand. On the other hand, it is narrow for low output power demand.

Conclusions
In this paper, the basic knowledge and operation of an SRG and its drive have been provided. The SRG was operated only under single-pulse-mode operation since this mode will have higher efficiency. The SRG operates at high speeds. Nevertheless, in this study, we neglected the output power profiles at low and medium speeds. The phase current and dc-link voltage waveforms were measured from the highly accurate current sensors and voltage sensor circuits. Output power control was implemented by integrating the ANN. The ANN was trained to effectively provide the optimal turn-off and turn-on angles for a certain output power profile. The simulation and experimental results were similar, showing that the proposed method is valid and practical.