Inter national J our nal of P o wer Electr onics and Dri v e System (IJPEDS) V ol. 16, No. 4, December 2025, pp. 2197 ∼ 2211 ISSN: 2088-8694, DOI: 10.11591/ijpeds.v16.i4.pp2197-2211 ❒ 2197 Speed contr ol of 3-phase induction motor with modied DTC using HT AF-ANN Ar pita Banik 1 , Raja Gandhi 2 , Chandan K umar 3 , Ach yuta Nand Mishra 3 , Rak esh Roy 1 1 Department of Electrical Engineering, National Institute of T echnology Me ghalaya, Shillong, India 2 Department of Electrical Engineering, Supaul Colle ge of Engineering Supaul, Bihar , India 3 Department of Electronics and Communication Engineering, Supaul Colle ge of Engineering Supaul, Bihar , India Article Inf o Article history: Recei v ed Mar 13, 2025 Re vised Jul 26, 2025 Accepted Sep 2, 2025 K eyw ords: Articial neural netw ork Direct eld-oriented control Direct torque control Hyperbolic tangent acti v ation function Induction motor Mean square error PI controller ABSTRA CT In this research paper , an art icial neural netw ork (ANN) algorithm is implemented with modi cations to enhance the performance of a direct torque controlled (DTC) induction motor dri v e. Since the main challenge in the con v entional DTC technique is to tune the PI controller appropriately therefore in this w ork, an ANN technique is incorporated in place of the con v entional PI controller . Sudden changes in speed and loading in induction motor dri v es lead to sharp uctuations and disturb the motor performance. In order to o v ercome these issues, a trained ANN controller is initially used here to enhance motor dri v e performance. Subsequently , the performance is further impro v ed by modifying the ac ti v ation function in the ANN controller . Here, motor parameters at rated and v ariable speed with v arious loading conditions ha v e been analyzed and compared for the DTC with a c on v ent ional PI controller with ANN, and a proposed ANN contr oller . Simulation of the complete model with the con v entional and proposed controllers is done using MA TLAB/Simulink platform to observ e the v arious speed responses for dif ferent conditions, and the e xperimental setup is used to demonstrate the ef fecti v eness and performance of the proposed system. This is an open access article under the CC BY -SA license . Corresponding A uthor: Arpita Banik Department of Electrical Engineering, National Institute of T echnology Me ghalaya Shillong, Me ghalaya, India Email: p19ee011@nitm.ac.in 1. INTR ODUCTION In recent years, induction motors (IMs) ha v e replaced DC machines in industrial dri v e applicat ions with its properties lik e lo w cost, easy maintenance, and rob ust structure [1], [2]. But the non-linear beha viour of IM necessitates v arious types of control strate gies to be implemented for obtaining better performance from the motor . V arious researchers ha v e disco v ered the f act that IM performs better with the v ector control technique than scalar control, which are the tw o major control techniques in control engineering [3]. In IM the major function of v ector control t heory is to determine the phase angle and magnitude of v oltages and currents. This v ector control method operated on the basis of P ark and Clark e transformations, which generates ux and torque of the motor , respecti v ely [4]. The tw o leading high-performance control techniques for IM dri v e in recent years are eld-oriented control (FOC) and direct torque control (DTC). The result in [5], [6] pro vide good dynamic and steady state torque responses, and also because of the ability of decoupl ing control between ux and torque. In the DFOC technique which w as proposed by Aziz et al. [7] and Blaschk e [8], tw o Hall ef fect sensors are J ournal homepage: http://ijpeds.iaescor e .com Evaluation Warning : The document was created with Spire.PDF for Python.
2198 ❒ ISSN: 2088-8694 used in the air g ap for determining rotor ux. The dra wback of the direct eld-oriented control (DFOC) method is that i t gi v es poor dynamic perform ance when the resistance v alue of the stator and rotor increases. Ease of use, lo w sensiti vity to parameter v ariations, rapid torque response, and simple implementation has made DTC a better choice than FOC. Moreo v er DTC approach doesn’ t require the coordinate transformation or the current re gulators [5]. This control technique, which w as introduced by T akahashi and Noguchi in Japan in the mid of 1980’ s, operates solely in a stationary frame (co-ordinates x ed to the stator) in contrast to FOC [9]. Also it eliminates the need for an y modulation, such as pulse width modulation (PWM), by producing the in v erter’ s g ating signals directly through the look-up switching table [10]. In DTC selection of in v erter switching sector for the re gulation of ux and torque is a v ery crucial part as a slight dif ference in t h e v ector during the v oltage sector selection process may cause a substantial phase shift mistak e in the command torque and cause ripples in torque and current [11]-[13]. When v ariable frequenc y operations are applied by h yst eresis comparators, con v entional DTC has dra wbacks such as decreased rob ustness due to v ariations in stator and rotor resistances. These elements shorten the machine’ s lifespan by raising the harmonics in the system, producing audible noises, and causing mechanical vibrations [14]. By inte grating optimization and machine learning techniques discussed in [15], the dra wbacks of con v entional DTC techniques ha v e been o v ercome, and thereby motor ef cienc y could be impro v ed. Use of these uni v ersal approximation methods through v arious controllers’ non-linearities of motors also ha v e been addressed in [16]. Study in [17] used a model reference adapti v e system (MRAS), which could impro v e the performance of the control scheme by gi ving f aster speed response and high rob ustness ag ainst e xternal load disturbance and reference speed v ariation. A research article [18] presents a model reference adapti v e system based on the acti v e po wer (P-MRAS) which can estimate stator resistance where the model is insensiti v e to parameter v ariations. Here a ne w structure of 12-sector DTC in combination with P-MRAS estimator and a back stepping speed controller is proposed to impro v e the direct torque control (DTC) strate gy . Mahfoud et al. [19] ha v e used genetic algorithm (GA) based DTC approach for optimizing K p , K I , and K D parameters and ha v e presented a comparison of dif ferent objecti v e functions with respect to speed o v ershoot and rejection time, ux es, and torque ripples and current total harmonic distortion (THD). Control technique lik e direct torque fuzzy control (DTFC), direct torque neuro control (DTNC), and direct torque neuro-fuzzy control (DTNFC) has impro v ed the ef cienc y of the con v entional technique and helped the system to perform better . Hysteresis comparators and truth tables ha v e been replaced in DTNFC by fuzzy logic and ANN. The torque and ux can be na vig ated to w ards their references o v er a predetermined amount of time because of the v oltage v ector created by ANFIS, which combines articial neural netw ork (ANN) and fuzzy logic. Using neural netw orks and fuzzy logic techniques, DTC dri v es’ PI speed controller is upgraded in [20]-[22]. In this research article, the authors ha v e discussed dif ferent con v entional controllers along with v arious machine learning techniques implemented in them for the performance enhancement of IM dri v e system. All the techniques discussed here has some adv antages and also some dra wbacks and limitations. T o mitig ate issues lik e high torque and ux ripples, reduced control accurac y at lo w speed, and dif culty in optimal tuning, machine learning techniques ha v e started been implemented in DTC of IM to mak e the motor performance better . ANN technique can model comple x and nonlinear motor dynam ics and also has the capability of learning control strate gies from training data. Once the system is trained ANN mak es the signal generation f aster . The accurac y of the stator ux location, which can be determined using the ANN technique, is crucial for sector v erication in the DTC scheme. The qual ity of reducing the dependenc y on identifying accurate motor parameters mak es ANN a better choice for controlling a non-linear system. Modern industry’ s requirement for intelligent motor dri v es has been met with techniques lik e ANN, as it made the s ystem sensor less and self-tuned. A study in [23] e xplains ho w ne-tuning the con v entional PI speed controll er wi th art icial intelligence can impro v e the upgraded DTC IM dri v e’ s performance. Aziz et al. [7] discussed about the use of the ANN technique in v arious industri al applications, and the re sults obtained from those studies pro v es the ability of the ANN technique in gi ving better results to w ards solving engineering problems than con v entional techniques. In this study , the performance of traditional DTC is enhanced through the application of the ANN technique. The result in [24] achie v ed an impro v ement of 85% in response time, reduction of speed o v ershoot, and 50% reduction in torque ripples by using the ANN technique in a doubly fed IM. Djeriri et al. [25] an ANN-based DTC controller is used for a doubly fed induction generator in selecting switching v olatge v ectors. Here ’ tansig’ and ’purelin’ functions are used for hidden and output layers, respecti v ely . Aissa et al. [26], a Int J Po w Elec & Dri Syst, V ol. 16, No. 4, December 2025: 2197–2211 Evaluation Warning : The document was created with Spire.PDF for Python.
Int J Po w Elec & Dri Syst ISSN: 2088-8694 ❒ 2199 neural switching table along with a fuzzy logic controller is used for brok en bar f ault diagnosis of IM. Here, also’ tansig’ and ’purelin’ functions are used as hidden and output layers, respecti v ely . In this research an ANN-based DTC model with sigmoid hidden neurons and tansig output neurons is de v eloped for the enhancement of the IM’ s transient performance under v aried dynami c situations. The Simulink model is rst run with the trained ANN controller and then the system responses are check ed with v arious ANN acti v ation functions lik e Sigmoid, h yperbolic tangent, rectied linear unit (ReLU), leak y ReLU, softmax. After running the simulation model it is observ ed that with si gmoid, ReLU, and Leak y ReLU functions, the speed response of the motor going be yond the rated speed, and with the Softmax function, the motor is running at belo w rated speed. During this process it has been observ ed that the model is gi ving better response with the ’h yperbolic tangent’ funct ion compared to the inb uilt function. Therefore, in this w ork the motor responses are v eried and compared for the con v entional systems with the proposed ANN model with the ne w function. The structure of this research article is or g anized as: i) Section 2 e xplains the system in general; ii) Sect ion 3 e xplains ANN structure and its algorithm; iii) Section 4 presents the methodology of proposed HT AF-ANN technique for DTC of IM; i v) Section 5 presents results and discussions; and v) This research article is concluded in section 6 which is follo wed by further research perspecti v e in future. 2. SYSTEM O VER VIEW 2.1. Con v entional DTC of IM In this system, the v alues of reference ux and speed are compared with the actual v alues. In the con v entional technique, the speed error through the PI controller generat es torque. The ux and torque error signal, after being controlled by a h ysteresis controller , is used to produce a re gulated switching signal so that the in v erter can function. The correct v ector is determined from the lookup database based on the torque error , ux v ector angle, and ux error . After the selected v ector is sent to the in v erter , the con v erter output po wers the motor [27]. The mathematical equations mainly used for designing DTC of 3-phase IM with respect to the stator reference frames are listed here as (1) and (2). ψ ds = Z ( V ds − R s i ds ) dt (1) ψ q s = Z ( V q s − R s i q s ) dt (2) Where, ψ ds , ψ q s represents stator d and q ax es ux linkages and V ds , V q s , i ds , i q s represents the stator d and q ax es v oltage and current components recepti v ely . Finally the electromagnetic torque of 3-phase IM can be e xpressed as (3). T e = 3 2 P 2 ( ψ ds i q s − ψ q s i ds ) (3) Where, P is the total no of poles in the IM. Con v entional DTC technique requires proper tuning of PI controller to achie v e e xpected output from the motor . V arious tuning methods are discussed in [28] among which ’ trial and error’ method is the most popular approach that decides g ain parameters of the controller using (4). u ( t ) = K p e ( t ) + K i Z t 0 e ( t ) dt (4) Where, u(t) represents controlled ou t put. K p stands for proportional g ain. K i is the inte gral g ain. e(t) stands for error signal and can be represented as the subtracti v e v alue of reference input and the process v ariable. But this method gi v es sluggish and improper response. Therefore optimization and machine learning techniques started getting implemented to o v ercome the shortcomings of PI controller . In this article DTC model of IM is trained with a proposed ANN model to get better performance of the machine dri v e system. 2.2. Pr oposed DTC with HT AF-ANN contr oller A model that emulates the structure and operation of actual neural netw orks found in animal br ains is called a neural netw ork (NN). It is basically a connection of articial neurons. A non-linear function called acti v ation function is used to determine the output of each neuron. W eight denes the connection from one Speed contr ol of 3-phase induction motor with modied DTC using HT AF-ANN (Arpita Banik) Evaluation Warning : The document was created with Spire.PDF for Python.
2200 ❒ ISSN: 2088-8694 neuron to another which determines the strength and direction of the inuence one neuron has on another . In an ANN, output is determined by a series of mathematical operations [29]. Figure 1 represents t he block diagram of DTC of 3-phase IM with proposed ANN controller . One neural netw ork is suggested in this paper to maximize the performance of the PI controller . Here, the speed error is estimated using a neural netw ork and to generate the electromagnetic torque. This neura l netw ork is associated with one input which is tak en by comparing actual with reference speed of the motor and one output. It is a f eed forw ard, tw o-layer netw ork with 10 neurons in hidden layer . Here Le v enber g Marquardt algorithm is used and the weights and bias v alues are updated according to this technique. After the t raining process is o v er if the output doesn’ t meet the e xact requirement then the training process is repeated. A neural netw ork’ s tendenc y is to approximate the output for ne w input data because of which the y are used in intelligent system analysis [30]. Figure 1. Block diagram of DTC of 3-phase IM with proposed ANN controller 3. ANN STR UCTURE AND ITS ALGORITHM 3.1. T raining data generation ANN technique can minimize the ef fort in e v aluating appropriate K p K i v alues in con v entional PI controller . An elaborate training process for ANN controller is discussed in [31]. In this w ork to replace PI controller an ANN is used to predict motor’ s reference torque T ∗ e taking speed error as input. Input matrix for the ANN controller is as: X = e ω ˙ e ω ω r T L Where, speed error e ω is the dif ference between the motor’ s actual speed and its rated speed, e . ω is the rate of change of speed error and T L is the load torque. Here the output is as: Y = T ∗ e T ∗ e is the optimal torque. Figure 2 sho ws the architecture of ANN. It has to predict the optimal torque by learning the mapping of speed error to the reference torque. T raining data is generated using a con v entional PI controller ensuring proper tuning of K p and K i . Simulation is run with v ariable speed reference and with step changes in load torque T L . Input and output v ariables are sa v ed in w orkspace in e xporting the data to MA TLAB. Once the data is collected ANN is trained using MA TLAB’ s NN T oolbox. Ultimately , a MA TLAB function block is Int J Po w Elec & Dri Syst, V ol. 16, No. 4, December 2025: 2197–2211 Evaluation Warning : The document was created with Spire.PDF for Python.
Int J Po w Elec & Dri Syst ISSN: 2088-8694 ❒ 2201 de v eloped and used in place of PI controller . At last simulation is run wi th ANN controller block and the motor response is analyzed. Figure 2. Architecture of ANN 3.2. T raining of the netw ork In MA TLAB/Simulink the process follo wed by a neural netw ork to solv e DTC of IM model can be e xplaine d as follo ws: In MA TLAB neural netw ork simulation generates a functi on as ’myNeuralNetw orkFunction’ which tak es an N × 1 matrix as input and gi v es an N × 1 matrix as output using the trained netw ork. This function is represented by the MA TLAB neural netw ork as in (5). f unctiony 1 = my N eur al N etw or k F unction ( X 1 ) (5) Where the input function is x 1 and output function is y 1 . Prior to data being fed into the neural netw ork, this stage in v olv es scaling the input data using normalization parameters. The equation used in this step is as (6). X p 1 = ( x 1 − x of f set ) .g ain + y min (6) Where, x of f set is the of fset v alue subtracted from the input data, g ain is the f actor by which input data is multipli ed, y min is the minimum v al ue after scali ng. In this step rst and second layer parameters are dened which includes weights and biases for both the layers as I W 1 1 & b 1 and LW 2 1 & b 2 respecti v ely . Computation of hidden layer is done for both the layers using (7) and (8) for layer 1 and layer 2 respecti v ely . a 1 = tansig ( W 1 X p 1 + b 1 ) (7) a 2 = L W 2 1 .a 1 + b 2 (8) In this step output normalization parameters are dened to get the output data back to its origi nal range. Equation follo wed in this step is as (9). y 1 = ( a 2 − y min ) /g ain + x of f set (9) Where, y min is the minimum v alue after scaling, g ain is the f actor by which output data is di vided, x of f set is the of fset v alue added to the output data. In this step simulation process starts where input data x 1 is transposed and normalized using mapminmax apply function. The output of the rst layer is calculated using ’ tansig apply’ function ha ving an e xpression of (10). f unctiona = tansig appl y ( n, ∼ ) a = 2 1 + e − 2 n − 1 (10) Then the output in the second layer is computed and the same is denormalized using ’mapminmax re v erse’ function and the nal output is then denormal ized to match the original data scale. In the proposed method the g ain f actor of (6) is changed and the output is computed using h yperbolic tangent function, e xpressed as (11). f unctiona = tanh appl y ( n, ∼ ) a = e αn − e − αn e αn + e − αn (11) α is a parameter here which scales the input. Speed contr ol of 3-phase induction motor with modied DTC using HT AF-ANN (Arpita Banik) Evaluation Warning : The document was created with Spire.PDF for Python.
2202 ❒ ISSN: 2088-8694 MSE is a loss functi on in ANN that is used to e v aluate the netw ork’ s performance. It is calculated by comparison between actual tar get v alues and the netw ork’ s predicted results. MSE can be e xpressed by (12). M S E = 1 N N X i =1 { y act,i − y pr ed,i } 2 (12) Where, N is no. of data points, y act,i and y pr ed,i are the actual tar get v alue and predicted output from the netw ork for i − th data point respecti v ely . The weight update e xpression, found in (13), is us ed to modify the weight of each neuron in order to lo wer the cost function and MSE v alues. T o update rule using gradient descent follo wing is an e xpression for the weight w ij that joins neurons i in the rst layer to neurons j in the subsequent layer , as (13). w ij ( t + 1) = w ij ( t ) − η ∂ M S E ( t ) ∂ w ij ( t ) (13) Where, t is the iteration, w ij is the weight, η is the learning rate, ∂ M S E ( t ) ∂ w ij ( t ) is the gradient of the MSE with respect to the weight w ij . In a NN, an acti v ation function is a mathematical function that is applied to a neuron’ s (or node’ s) output to cause the model to become non-linear . 4. METHODOLOGY OF PR OPOSED HT AF-ANN TECHNIQ UE FOR DTC OF IM As discussed in introduction section a neural netw ork may ha v e v arious acti v ation functions which can be chosen basis the problem statement. In this research after v erifying motor responses with all the functions it is found that h yperbolic tanget is gi ving the best results for this problem statement. So inb uilt function is replaced here by the proposed function. The e xpression for this h yperbolic tangent function is gi v en in (11). Simulation results sho ws that this proposed ANN algorithm has gi v en signicant impro v ement in motor responses compared to con v entional PI controller and ANN controller . Figure 3 sho ws the o wchart of proposed ANN control technique for DTC of 3-ph IM. The simulation steps follo wed for impro ving the technique is e xplained here in the char t. Simulation results obtained from the model follo wing these steps ha v e gi v en a signicance impro v ement in the motor performance and the results are listed in section 5. Figure 3. Flo wchart of proposed ANN technique Int J Po w Elec & Dri Syst, V ol. 16, No. 4, December 2025: 2197–2211 Evaluation Warning : The document was created with Spire.PDF for Python.
Int J Po w Elec & Dri Syst ISSN: 2088-8694 ❒ 2203 5. RESUL TS AND DISCUSSIONS Simulink model of DTC of 3-phase IM is de v eloped in MA TLAB platform here. Since the main tar get of this research is to control the speed and torque parameters of the motor , hence the trained ANN controller tak es place of the PI controller when it has been tuned. Here, the DTC analysis is conducted using a 750 V A IM. T able 1 contains a l ist of the motor parameters. The rated speed of the motor is 145.56 rad/sec. The o wchart in Fi g ur e 3 discusses the research process that is incorporated here. In this article rst the responses obtained during the ANN training is discussed and then the motor responses for three dif ferent dynamic cases are discussed. Motor parameters lik e t s , t r , t p , %Mp, speed and torque ripples are analyzed and compared here for the proposed technique with the con v entional one. Also e xperimental results with speed and torque v ariation has been included here. 5.1. P erf ormance analysis of ANN contr oller Figure 4(a) sho ws the operation of gradient training for induction motor speed at an epoch of 1000. It can be seen in the gure that the range of mean gradient v ariation is 10 0 to 10 5 . Figure 4(b) sho ws each test point’ s v alidation check. Here, it is e vident that e v ery sample passed the test o v er the course of 1000 epochs. Figure 4(c) displays the mean squared error performance across 1000 epochs. At epoch 1000, the optimal training performance is 0.0020364. The estimated and actual data re gression analysis is sho wn in Figure 5(a). It sho ws that for R = 0.99931, 0.99927, and 0.99914 the training, v alidation and testing is close to the trajectory . Finally with R = 0.99928 v alidates the complete model by e xactly meeting the trajectory path. The error histogram is sho wn in Figure 5(b). Here, the ANN model’ s entire error range is distinguished into 20 bins. T otal error in this research ranges from -6.155 (left side bin) to 2.446 (right side bin). T able 1. Specications of the IM Motor parameters v alue unit Nominal po wer 750 V A Frequenc y 50 Hz Stator resistance 8.6 Ohm Stator inductance 0.045 Henry Rotor resistance 6.26 Ohm Rotor inductance 0.045 Henry Mutual inductance 0.709 Henry Pole pairs 2 - Inertia 0.0002 kg.m-2 Figure 4. ANN training pattern analysis: (a) gradient training, (b) v alidation check, and (c) mean square error performance across training, testing and v alidation 5.2. Assessment of the simulation r esults The implemented model of ANN based DTC of IM using MA TLAB/Simulink is sho wn in Figure 6. At rst DTC algorithm for IM is de v eloped in MA TLAB using Simulink blocks. Here ux and torque Speed contr ol of 3-phase induction motor with modied DTC using HT AF-ANN (Arpita Banik) Evaluation Warning : The document was created with Spire.PDF for Python.
2204 ❒ ISSN: 2088-8694 estimation blocks are designed using the measured v alues of stator v oltages and currents. T o get the in v erter switching pulses switching table logic is de v eloped and nally the signal is sent to the IM. In this research rst con v entional DTC of IM using a PI controller is modeled and later the ANN technique is incorporated to replace the PI controller for impro ving the performance of the system. The system is further modied by implementing the proposed technique and the obtained results are compared with the con v entional technique responses. Here the performance analysis is done for v arious speed and load conditions of the motor using MA TLAB/Simulink. Figure 5. Results of the re gression analysis: (a) re gression analysis for ANN and (b) histogram for re gression analysis Figure 6. Simulink model of ANN based DTC Int J Po w Elec & Dri Syst, V ol. 16, No. 4, December 2025: 2197–2211 Evaluation Warning : The document was created with Spire.PDF for Python.
Int J Po w Elec & Dri Syst ISSN: 2088-8694 ❒ 2205 5.2.1. Speed analysis This section discusses the results of the MA TLAB simulation and e v aluates the dif ferent motor parameters based on the responses. In this instance, the entire analysis is di vided into three cases, as follo ws: − Case 1: In v estig ation of motor response during starting at rated speed under no load condition: In this particular case, the motor is running at its rated speed of 145.56 rad/sec with no load. At rs t the PI controller is tuned with suitable K p and K i v alues and the DTC model of IM is run at rated speed. Then with the data collected in w orkspace ANN is b uilt and the model is run with the trained ANN controller and nally the model is run with proposed ANN controller and the responses are collected and compared. Motor transient responses at rated speed (145.56 rad/sec) under no-load with v arious controllers are sho wn in Figures 7(a), 7(b), and 7(c), and a list of e v ery parameter that w as e xamined from these w a v eforms is pro vided in T able 2. It i s e vident from the v alues of e v ery parameter gi v en in T able 2 that the proposed controller performs signicantly better than the traditional PI and ANN for case 1. − Case 2: Examination of motor response and parameters with v ariable speed command under no load condition with PI, ANN and proposed controller: Figures 8(a), 8(b), and 8(c) sho ws the speed responses of DTC of IM model with all three controll ers at 25% (36.39 rad/sec), 50% (72.78 rad/sec), 75% (109.17 rad/sec), and 100% (145.56 rad/sec) of rated speed. The motor parameters for all these four speed commands are listed in T ables 3, 4, and 5 respecti v ely . From the v alues tab ulated here it is e vident that the proposed controller is making the motor perform better compared to the other tw o controllers. Here all the responses are meticulously e xamined to ensure t he better performance of the motor . Figure 7. Speed responses at rated speed under no-load with (a) PI controller , (b) ANN controller , and (c) proposed ANN controller T able 2. Motor parameters with dif ferent controllers under no load and at rated speed P arameters PI C on v enti onal ANN Impro v ed ANN Impro v ement Impro v ement w .r .t w .r .t PI (%) con v enti onal ANN (%) %Mp 5.33 4.58 3.04 42.96 33.62 t r (sec) 0.0112 0.0102 0.01 10.71 1.96 t p (sec) 0.019 0.018 0.0168 11.58 6.67 t s (sec) 0.08 0.07 0.03 62.5 57.14 Speed ripple (rad/sec) 0.081 0.07 0.062 23.46 11.43 Speed contr ol of 3-phase induction motor with modied DTC using HT AF-ANN (Arpita Banik) Evaluation Warning : The document was created with Spire.PDF for Python.
2206 ❒ ISSN: 2088-8694 Figure 8. Speed responses with v ariable speed commands under no-load with (a) PI controller , (b) ANN controller , and (c) proposed ANN controller T able 3. Motor parameters with PI controller under no load and at v ariable speed P arameters 25% of rated speed (36.39 rad/sec) 50% of rated speed (72.78 rad/sec) 75% of rated speed (109.17 rad/sec) 100% of rated speed (145.56 rad/sec) %Mp 18.16 17.10 8.21 4.36 t r (sec) 0.003 0.0013 0.0011 0.0014 t p (sec) 0.0087 4.0025 8.0025 8.0025 t s (sec) 0.2 4.007 8.0055 8.0055 Speed ripple (rad/sec) 0.02 0.045 0.039 0.039 T able 4. Motor parameters with ANN controller under no load and at v ariable speed P arameters 25% of rated speed (36.39 rad/sec) 50% of rated speed (72.78 rad/sec) 75% of rated speed (109.17 rad/sec) 100% of rated speed (145.56 rad/sec) %Mp 15.14 14.32 6.256 3.05 t r (sec) 0.0028 0.0013 0.0011 0.0014 t p (sec) 0.0085 4.0025 8.0025 12.0025 t s (sec) 0.015 4.007 8.0055 12.007 Speed ripple (rad/sec) 0.019 0.045 0.039 0.039 T able 5. Motor parameters with proposed ANN controller under no load and at v ariable speed P arameters 25% of rated speed (36.39 rad/sec) 50% of rated speed (72.78 rad/sec) 75% of rated speed (109.17 rad/sec) 100% of rated speed (145.56 rad/sec) %Mp 9.92 8.54 4.42 2.02 t r (sec) 0.002 0.0012 0.00129 0.001 t p (sec) 0.008 4.002 8.0023 12.0025 t s (sec) 0.008 4.006 8.0052 12.005 Speed ripple (rad/sec) 0.019 0.046 0.039 0.062 − Case 3: Ev aluation of motor response at rated speed with v ariable load conditions In this case the motor beha viour is e xamined for sudden change in load torque. At rst, the motor operates at its rated speed of 145.56 rad/sec without an y load, and then motor is suddenly e xposed to a 5.15 N-m load. The per formance of the motor is observ ed here precisely as t his case conditions determines the rob ustness of the motor and ensures the ef fecti v eness of the system in managing dynamic load changes. The performance of the motor under dif ferent load conditions with dif ferent controllers are sho wn in Figures 9(a), 9(b), and 9(c). The v arious motor parameters related to speed response under no load condition is gi v en already in T able 2 and same for full load is listed in T able 6. Int J Po w Elec & Dri Syst, V ol. 16, No. 4, December 2025: 2197–2211 Evaluation Warning : The document was created with Spire.PDF for Python.