Application of Artificial Neural Networ in Economic Generation Scheduling of Thermal ower lants Mohammad Mohatram Department of Electrical & Electronics Engineering Sanjay Kumar Department of Computer Science Engineering Waljat Colleges of Applied Sciences, Muscat, Sultanate of Oman, with Birla Institute of Technology, Mesra (Ranchi), India Email: mohatram@waljatcolleges.edu.om, mmohtram20@rediffmail.com Abstract- The artificial neural networ has been applied to the economic generation scheduling of six thermal power plants with very promising results. In this problem equality constraints of power balance and inequality plant generation capacity constraints are considered. The transmission losses occurring are also considered. The inputs to the neural networ contain the total load supplied. The electric power generation of six thermal power plants and the total system transmission losses are taen as the output of the neural networ. A program in C language is developed to generate training and test pattern for the networ developed. At the end performance and time taen in execution of the neural networ is compared with the Classical Kirchmayer Method and it is observed that the neural networ being very fast and accurate. Therefore, it may replace effectively the conventional practices presently performed in different central load dispatch centers.. Introduction The size of electric power system is increasing rapidly to meet the energy requirements. A number of power plants are connected in parallel to supply the system load by interconnection of power stations. With the development of grid system it becomes necessary to operate the plant unit most economically. The economic generation scheduling problem involves two separate steps namely the unit commitment and the online economic dispatch. The unit commitment is the selection of unit that will supply the anticipated load of the system over a required period of time at minimum cost as well as provide a specified margin of the operating reserve. The function of the online economic dispatch is to distribute the load among the generating units actually paralleled with the system in such a manner as to minimize the total cost of supplying the
minute to minute requirements of the system [, 4]. Thus, economic load dispatch problem is the solution of a large number of load flow problems and choosing the one which is optimal in the sense that it needs minimum cost of electric power generation. Accounting for transmission losses results in considerable operating economy. Further more this consideration is equally important in future system planning and in particular, with regard to the location of plants and building of new transmission lines. To calculate electric power generation of various units with different load demands (Training data), the usual Classical Kirchmayer Method is used. In this paper an Artificial Neural Networ [] based method is proposed. can be defined as a class of mathematical algorithms designed to solve a specific problem. Basically s are parallel computational models comprised of densely interconnected adaptive processing units. An extremely important and human characteristic of s is their adaptive nature, where learning by experience replaces programming in solving problems [2, 3]. s learn the pattern on which they are trained. In this wor is trained with different load demands. Once it has been trained, it acquires the ability to give load scheduling pattern for any value for load demand. 2. roblem Formulation A number of thermal power plants are connected to a common grid which supplies power to different load centers. The load demand is totally at the discretion of the consumers and it varies over a wide range. The cost of power generation is not the same for every unit. So, to have the minimum cost of generation for a particular load demand, we have to distribute the load among the units which minimize the overall generation cost with the constraint that no unit is overloaded. The majority of generating units have a non linear cost function C i. The variation of fuel cost of each generator with active power output Gi is given by a quadratic polynomial. C i a i 2 Gi + b i Gi + d i () Where a i is a measure of losses in the system, b i is the fuel cost and d i is the salary and wages, interest and depreciation. The optimal dispatches for the thermal power plants should be such that the total electric power generation equal to the load demand plus line losses, which can be written as: K i Gi D L 0 (2) where, K total number of generating plants, 2
Gi generation of i th plant, L total system transmission loss, D system load demand The transmission losses which occur in the line when power is transferred from the generating station to the load centers increases with increase in distance between the two [2, 4, 6]. The transmission losses may vary from 5 to 5 % of the total load. If the power factor of load at each bus is assumed to remain constant the system loss L can be shown to be a function of active power generation at each plants i.e. L L ( G, G2 GK ) (3) One of the most important, simple but approximate method of expressing transmission loss as a function of generator powers is through B-Coefficients and is given by Kron s loss formula [] as: L B + Gi ij Gj i j i B0 i Gi + B00 (4) Where, Gi, Gj real power generation at i th and j th power unit B ij, B 0i loss coefficients, constant for certain conditions B 00 loss constant The inequality constraints is given by Gmin Gi Gmax (5) The maximum active power generation Gmax of source is limited by thermal consideration and minimum active power generation Gmin is limited by the flame instability of a boiler [, 4, 6]. 3. Software Description A program in C language is developed for generating training patterns by Classical Kirchmayer Method and Error Bacpropagation Method. 3. Classical Kirchmayer Method: This method [, 4, 6] is used to generate the output data for six thermal power plants considering transmission losses. The Algorithm is as follows:. Start 2. Read the constants a i, b i, loss coefficient matrices B ij, and B 0i, constant B 00, power demands D, maximum GMX, minimum GMN generators real power limits 3. Assume a suitable value of λ λ 0. This value should be greater than the largest intercept of the incremental cost of the various units. Calculate G, G2, -----------, Gi based on equal incremental cost. 4. Calculate the generation at all buses using B bi λ 2ai + 2B λ 2B 0i ij Gj j j Gi, ii i,2, (6) 3
eeping in mind that the values of powers to be substituted on the RHS of Eq.(6) during zeroth iteration correspond to the values calculated in step 3. For subsequent iterations the values of powers to be substituted corresponds to the powers in the previous iteration. However if any generator violates the limit of generation that generator is fixed at the limit violated. 5. Chec if the difference in power at all generator buses between two consecutive iterations is less than the specified value, otherwise go bac to step 2 6. Calculate the losses using the relation L B + Gi ij Gj i j i B 0 i Gi + B (7) 7. Calculate, i Gi D L 00 (8) 8. If is less than a specified value ε, stop calculation and calculate cost of generation with values of powers. If < ε is not satisfied go to step7. 9. Update λ as λ (+) λ () λ () where λ is the step size [7, 8, 9 ]. 0. Stop 3.2 Bac ropagation Algorithm:. Start 2. Read the input values, (X), target values, (T), learning rate coefficient, (ή), tolerance limit of error, (ε), momentum constant, (α), maximum number of iterations, (itrmax), neurons in hidden layer, (n), input nodes, (m), output nodes,(z), total number of input pairs, (p), maximum value of input, (x_max), maximum value of target, (t_max), and constant, (SEED) to generate weight matrices. 3. Normalize the input and target vector. 4. Activate the first node of the input by.0 and the rest by the normalized inputs. 5. Generate the random weights. 6. Set all the elements of error weight matrices dw and dw2 equal to zero. 7. Set i 0.0 8. Set serr 0.0 9. Calculate output Y using Y (i, j) X (i, ) * W(, j) where j0,, n-; 0,, m-; 0. rocess output of hidden layer through a non linear activation function (positive sigmoid function) to produce output of hidden layers. Y (i, j) /(+exp(-y (i, j))). Calculate the final output using Y (i, ) Y (i, j) * W2 (j, ) 4
2. rocess again through sigmoid function to find output of networ. Y (i, ) /(+exp(-y (i, ))) 3. Calculate square error as e(i, ) T(i, ) Y (i,) serr + e(i, ) * e(i, ) T 4. Adjust the weights of output layer d2 (i, ) Y (i,) *(- Y (i,)) * e (i,) T dw2(j, ) ή*d2 (i,) * ( Y (i,j)) +α * dw2 (j,) W2 (j,) +dw2(j,) 5. Adjust the weight of hidden layer. sum 0.0; sum + d2(i,) * W2 (j,) d (i, j) sum * Y (i,j) *(- Y (i,j)) dw(, j) ή *d (i,j) * X (i, ) + α * dw (, j) W (, j) + dw(, j) 6. Increase by one and if i<p go to step 8. 7. Chec for tolerance, if serr > ε then go to step 7. 8. Stop 4. Application Example The cost functions for six thermal power plants [5] are shown in table. Table. Cost of i th a i b i d i unit C i 0 7.0 240 2 0.0095 0.0 200 3 0.0090 8.5 220 4 0.0090.0 200 5 0.0080 0.5 220 6 5 2.0 90 The inequality plant capacity constraints i.e. the generator s real power limits are given in table.2 Table.2 Generator Real ower Limits Generator Min. Max. 00 500 2 50 200 3 80 300 4 50 50 5 50 200 6 50 20 The loss coefficient matrix used is 0.002 0.0007 B ij 0.000 0.0005 0.0002 0.002 0.004 0.0009 0.000 0.000 0.0007 0.0009 0.003 0.0000 0.000 0.000 0.000 0.0000 0.0024 0.0008 0.0005 0.000 0.029 0.0002 [ 0.3908 0.297 0.7047 0.059 0.26 0.6635 ] B 0i and constants B 00 0.0056 4. Results The results are shown in Table.3 and.4 0.0002 0.000 0.0008 0.0002 0.050 (Appendix). It is to be noted that the results obtained by are very close to the results obtained by conventional classical method. 5. Conclusion This wor aim to carry out development for based method to determine economic generation scheduling considering transmission losses of thermal power plants 5
very efficiently and accurately. In a power system there is a large variation in load from time to time and it is not possible to have the load scheduling pattern for every possible load demand. As there is no general procedure for finding out the economical load scheduling pattern. This is where plays an important role as we need small number of training data sets for the training of. A trained can then be applied to find out the economical load scheduling pattern for a particular load demand in a fraction of second. There is no need of having loss matrices, B ij, B i0, and loss constant B 00 in an based method. Results obtained are very much closed to the results obtained by Kirchmayer Method s. A remarable saving in the computation time has been observed. Due to flexibility in several other practical constraints can also be easily incorporated as input-output information of the training sets. For future wor it is suggested to design a general for such problems. Also, methods can be thought of which reduced the training time. The effect of complexity of the neural networ on the performance of system may also be studied. References [] Hussain Ashfaq, Electrical ower Systems, CBS ublications, New Delhi, 994. [2] H. Hassoun Mohamad, Fundamentals of Artificial Neural Networs, HI ublication, New Delhi, 998. [3] O. Mohammad, D. ar, R. Merchant, T. Dinh, C. Tong, A. Azeem, J. Farah, K. Cheung, ractical Experiences with an Adaptive Neural Networ Short-term Load Forcasting System, IEEE Transactions on ower Systems, vol. 0, no., Feb. 995. [4] Wadhwa C. L., Electrical ower Systems, NAI ublications, New Delhi, 995. [5] Sadat Hadi, ower System Analyses, Mc Graw Hill ublication, New Delhi, 997. [6] Nagrath I. J. and Kothari D.., ower System Engineering, TMH ublications, New Delhi, 994. [7] Carpenter G. A. and Grossberg S., Neural Dynamics of Category Learning and Recognition; Attention Memory Consolidation and Amnesia, AAAS Symposium Series, 986. [8] Kusic George L., Computer-Aided ower Systems Analysis HI ublication, New Delhi, 998. [9] J. H. ar, Y. S. Kim, I. K. Eom, and K. Y. Lee, Economic Load Dispatch for iecewise Quadratic Cost Function Using Hopfield Neural Networs IEEE Transactions on ower System 6
Apparatus and Systems, vol. 8, no.3, pp. 030-038, Feb. 993. [0] Kirchmayer L. K., Economic Operation of ower Systems, John Willey and Sons, New Yor, 958. 7
8 Appendix %EL -0.0 0.002 0.03 0 0.002 0.00-0.006 0.004-0.002 0.004 LA 34.03 34.68 36.35 38.07 4.68 45.5 49.53 53.8 58.36 60.77 63.29 65.93 LOSSES () LK 33.94 34.69 36.46 38.04 4.69 45.52 49.54 53.73 58.4 60.73 63.26 65.98 %E6 0.084-0.03 0 0.038 0.02-0.02-0.057-0.009 0.05 0.22-0.085 A6 49.3 50.93 55.8 59.6 68.97 78.78 88.75 98.6 08. 2.6 6.9 2 K6 50 50.06 55.7 59.58 69.33 78.98 88.53 97.98 08 2.8 8.4 20 %E5-0.083-0.06 0.037 0.03 0.039 0.038 0.09-0.07-0.024-0.05-0.036 0.068 A5 05. 06.6 0.6 4.7 23 3.3 39.8 48.5 57.2 6.7 66.3 70.9 K5 04.4 06.5 4.8 23.3 3.7 40 48.3 57 6. 65.5 7.8 %E4-0.002 0.059 0.067-0.048-0.084-0.09-0.05 0.095 0.75 0.075-0.77 A4 87.6 88.99 93.62 98.3 07.8 7. 26. 34.5 42.3 45.8 49. 52.2 K4 87.5 89.49 94.2 98.27 07.3 6.3 25.2 34 43.4 47.9 50 50 %E3-0.064-0.002 0.038 0.08 0.027 0.02-0.0-0.03-0.046-0.039-0.062 A3 24.5 26.08 220.93 225.8 235.59 245.45 255.42 265.6 276.5 28.6 287.2 293 K3 23.6 26. 22.3 225.8 235.8 245.7 255.6 265.5 276 28. 286.8 293.8 %E2-0.064-0.0 0.03 0.027 0.029 0.05-0.03-0.07-0.039-0.03 0.053 A2 2.4 6 9.5 26.8 34. 4.5 49 56.7 60.6 64.6 68.8 K2 0.5 2.3 6.2 9.6 27 34.4 4.7 48.9 56.5 60.2 64.3 69.4 %E -0.046-0.00 0.034-0.0 0.02 0.023 0.08-0.0-0.03-0.045-0.03 0.059 A 307.42 309.36 34.22 39.06 328.72 338.37 348. 358 368.2 373.46 378.84 384.34 K 307.03 309.35 34.5 38.96 328.83 338.6 348.29 357.89 368.06 372.94 378.48 385.06 LOAD D 840 850 875 900 950 000 050 00 50 75 200 225 RESULT FOR MEAN SQUARE OR 4.466038 E -06 TABLE.3 RESULTS IN TRAINING MODE (COMARISON OF TEST RESULTS OBTAINED BY CHMAYER AND METHOD ECONOMIC LOAD SHARING INCLUDING LOSSES OF SIX THERMAL OWER LANTS) %EL 0.002 0.006 0.006-0.005-0.005-0.009-0.07 0.07 0.022 LA 33.7 35.24 37.38 39.85 43.57 47.49 5.64 56.04 60.77 64.33 65.39 66.48 LOSSES () LK 33.72 35.39 37.42 39.77 43.52 47.44 5.54 56.2 60.57 64.23 65.6 66.74 %E6 0.8-0.058 0.008 0.009 0.07-0.05-0.057-0.023-0.02 0.6-0.09-0.5 A6 48.49 52.6 57.82 64.22 73.83 83.76 93.7 03.4 2.59 8.59 20.23 2.84 K6 50 52. 57.89 64.3 74 83.6 93. 03.5 2.45 20 20 20 %E5-0.063 0.03 0.04 0.026 0.07-0.0-0.005-0.074-0.028 0.076 0.29 A5 04.28 08.23 3.06 8.8 27.3 35.58 44.3 52.82 6.7 68.3 70 7.89 K5 03.75 08.29 3.33 8.92 27.38 35.75 44.02 52.76 60.85 67.8 70.93 73.47 %E4 0.027 0.063 0.03-0.043-0.084-0.06-0.09 0.029 0.5-0.03-0.29-0.224 A4 86.26 90.84 96.43 03.03 2.46 2.65 30.4 38.5 45.8 50.36 5.57 52.75 K4 86.48 9.38 96.7 02.64.63 20.56 29.43 38.83 47.56 50 50 50 %E3-0.033 0.03 0.02-0.0 0.009 0.009-0.075-0.036 0.076 0.29 A3 23.9 28.02 223.85 230.69 240.5 250.42 260.48 270.83 28.6 289.5 29.83 294.9 K3 22.9 28.5 224.04 230.6 240.58 250.5 260.4 270.93 280.73 289.08 292.76 295.78 %E2-0.044 0.008 0.023 0.006 0.07 0.0-0.00-0.06-0.024 0.062 0.04 A2 0.34 3.84 8.09 23.3 30.4 37.78 45.23 52.82 60.62 66.28 67.93 69.6 K2 09.97 3.9 8.3 23.8 30.57 37.89 45.4 52.8 59.92 65.99 68.69 70.88 %E -0.009 0.03 0.05-0.06 0.002 0.008-0.072-0.025 0.074 0.9 A 306.45 3.3 37.2 323.89 333.54 343.22 353.02 363.06 373.46 38.02 383.23 385.45 K 306.37 3.42 37.25 323.74 333.56 343.39 352.94 363.5 372.6 380.73 384.3 386.92 LOAD D 835 860 890 925 975 025 075 25 75 20 220 230 TABLE.4 RESULTS IN NON-TRAINING MODE (COMARISON OF TEST RESULTS OBTAINED BY CHMAYER AND METHOD ECONOMIC LOAD SHARING INCLUDING LOSSES OF SIX THERMAL OWER LANTS)
9