diff options
Diffstat (limited to '1445/CH2/EX2.19')
-rw-r--r-- | 1445/CH2/EX2.19/Ex2_19.sce | 59 |
1 files changed, 30 insertions, 29 deletions
diff --git a/1445/CH2/EX2.19/Ex2_19.sce b/1445/CH2/EX2.19/Ex2_19.sce index 32395bf13..015f1ab16 100644 --- a/1445/CH2/EX2.19/Ex2_19.sce +++ b/1445/CH2/EX2.19/Ex2_19.sce @@ -1,50 +1,51 @@ //CHAPTER 2- STEADY-STATE ANALYSIS OF SINGLE-PHASE A.C. CIRCUIT //Example 19 +clc; disp("CHAPTER 2"); disp("EXAMPLE 19"); //VARIABLE INITIALIZATION -z1=4+(%i*3); //impedance in rectangular form in Ohms -z2=6-(%i*8); //impedance in rectangular form in Ohms -z3=1.6+(%i*7.2); //impedance in rectangular form in Ohms -v=100 //in volts +z1=4+(%i*3); //impedance in rectangular form in Ohms +z2=6-(%i*8); //impedance in rectangular form in Ohms +z3=1.6+(%i*7.2); //impedance in rectangular form in Ohms +v=100 //in volts //SOLUTION -//solution (i) -//Admittance of each parallel branch Y1 and Y2 +//SOLUTION (i) + +//Y1 and Y2 are admittances of each parallel branch Y1=1/z1; Y2=1/z2; disp("SOLUTION (i)"); -disp(sprintf("Admittance parallel branch 1 is %3.3f %3.3fj S", real(Y1), imag(Y1))); -disp(sprintf("Admittance parallel branch 2 is %3.3f+%3.3fj S", real(Y2), imag(Y2))); +disp(sprintf("Admittance parallel branch 1 is %3f %3fj S", real(Y1), imag(Y1))); +disp(sprintf("Admittance parallel branch 2 is %3f+%3fj S", real(Y2), imag(Y2))); disp(" "); -//solution (ii) -//Total circuit impedance Z=(Z1||Z2)+Z3 -z=z3+(z2*z1)/(z1+z2) +//SOLUTION (ii) + +z=z3+(z2*z1)/(z1+z2) //series and parallel combination of impedances disp("SOLUTION (ii)"); -disp(sprintf("Total circuit impedance is %3.3f %3.3fj S", real(z), imag(z))); -//solution in the book is wrong as there is a total mistake in imaginery part 7.2+0.798=11.598 -// -//solution (iii) -//Supply current I=V/Z -i=v/z; -function [z,angle]=rect2pol(x,y); -z0=sqrt((x^2)+(y^2)); //z is impedance & the resultant of x and y -angle=atan(y/x)*(180/%pi); //to convert the angle from radians to degrees +disp(sprintf("Total circuit impedance is %3f %3fj S", real(z), imag(z))); +//solution given in the book is wrong as j(7.2+0.798) cannot be equal to j11.598 + +//SOLUTION (iii) + +I=v/z; +function [Z,angle]=rect2pol(x,y); //function 'rect2pol()' converts impedance in rectangular form to polar form +Z=sqrt((x^2)+(y^2)); //z is impedance & the resultant of x and y +angle=atan(y/x)*(180/%pi); //to convert the angle from radians to degrees endfunction; -[z, angle]=rect2pol(real(i), imag(i)); +[Z, angle]=rect2pol(real(I), imag(I)); //disp(sprintf("%f, %f",z,angle)); -//disp(sprintf("%f, %f",real(i), imag(i))); +//disp(sprintf("%f, %f",real(I), imag(I))); pf=cos(angle*%pi/180); - disp("SOLUTION (iii)"); -disp(sprintf("The power factor is %4.2f",pf)); -//solution (iv) -//Power supplied by source = VI cosĪ¦ or I^2 . R -P=v*real(i)*pf; +disp(sprintf("The power factor is %f",pf)); + +//SOLUTION (iv) +P=v*real(I)*pf; //power supplied by source is either (VI cosĪ¦) or (I^2 . R) disp("SOLUTION (iv)"); -disp(sprintf("The power supplied by source is %d watt",P)); -//END +disp(sprintf("The power supplied by source is %f watt",P)); +//END
\ No newline at end of file |