MULTI-RESERVOIR OPERATION BY ADAPTIVE PHEROMONE RE-INITIATED ANT COLONY OPTIMIZATION ALGORITHM

Through a collection of cooperative agents called ants, the near optimal solution to the multi-reservoir operation problem may be effectively achieved employing Ant Colony Optimization Algorithms (ACOAs). The problem is approached by considering a finite operating horizon, classifying the possible releases from the reservoir(s) intopre-determined intervals, and projecting the problem on a graph. By defining an optimality criterion, the combination of desirable releases from the reservoirs or operating policy is determined. To minimize the possibility of premature convergence to a local optimum, a combination of Pheromone Re-Initiation (PRI) and Partial Path Replacement (PPR) mechanisms are presented and their effects have been tested in a benchmark, nonlinear, and multimodal mathematical function. The finalized model is then applied to develop an optimum operatingpolicy for a single reservoir and a benchmark four-reservoir operation problem. Integration of these mechanisms improves the final result, as well as initial and final rate of convergence. In the benchmark Ackley function minimization problem, after 410 iterations, PRI mechanism improved the final solution by 97 percent and the combination of PRI and PPR mechanisms reduced final result to global optimum. As expected in the single-reservoir problem, with a continuous search space, a nonlinear programming (NLP) approach performed better than ACOAs employing a discretized search space on the decision variable (reservoir release). As the complexity of the system increases, the definition of an appropriate heuristic function becomes more and more difficult; this may provide wrong initial sight or vision to the ants. By assigning a minimum weight to the exploitation term in a transition rule, the best result is obtained. In a benchmark 4-reservoirproblem, a very low standard deviation is achieved for 10 different runs and it is considered as an indication of low diversity of the results. In 2 out of 10 runs, the global optimal solution is obtained, where in the other 8 runs; results are as close as 99.8 percent of the global solution. Results and execution time compare well with those of well developed genetic algorithms (GAs).