NETWORK LINEAR-PROGRAMMING CODES FOR WATER-SUPPLY HEADWORKS MODELING

There is a class of water-supply headworks models that uses network li near programming to assign water within the network. The choice of alg orithm and computer code to solve the network linear programs formulat ed by such models becomes important when doing planning studies that m ay involve thousands of replicates of future stream-flow and demand co nditions. The computational performance of two public-domain FORTRAN c odes, NETFLO and RELAX, representing efficient implementations of the primal-simplex and primal-dual algorithms, were compared for two headw orks systems. The results showed that for the two small network linear programs typical of headworks modeling, the RELAX code used without p rior flow information was about 100% faster than the NETFLO code. When prior flow information was employed, further substantial gains in spe ed were realized. The second and subsequent iterations within a time s tep were solved in about one-third of the time for the first iteration . The actual overall speed gain depends on whether the prior informati on is used across time steps or just between the iterations in a time step.