One of the most important factors in utilizing the water resources is the quality management of the reservoirs. Lack of decisiveness and certainty in optimization of reservoirs operation, increases the complexity of the subject. In this research, using a fuzzy approach in optimization of the reservoir release, a real time operating model has been presented for Zayandeh-Rud dam. To achieve this goal, the Harmony Search (HS) algorithm integrated with fuzzy approach and the data based procedure of the Support Vector Machines (SVM) is used. The harmony search algorithm, as a robust evolutionary method, and SVM as a machine learning method proved to be very useful in reservoir operation. Reservoir operation in terms of forecasting the annual release policy is determined in a two steps process. First, the fuzzy optimization procedure determines the optimal release values based on historical inflows into the reservoir. In the second step, a data-based regression model will be trained to forecast the future release policies. For fuzzification of the parameters and objective function, membership functions have been used according to the terms and conditions governing the planning conditions. The results show that the model in the years with sufficient water, in addition to increasing the percentage of providing the demands, initiates a potential for storing the water and increases the volume of storage rate. In the years with insufficient water, the model will suggest utilizing this potential for compensating the shortage of water and balancing the water needs. Throughout the period of 55 years, the average percentage of providing water with respect to the demands is %75 per month. The model of fuzzy harmony algorithm, comparing to crisp harmony algorithm, has improved the objective function by %6. Development of SVM model, by using the values of optimized release found in the first step, allows the system to be trained based on rainfall and inflow to the reservoir from previous months in order to forecast the reservoir outflow for future periods. The effective parameters for training the suggested model, include inflow runoff, monthly demands and initial storage volume. The proximity of the measured errors of the model, in training and testing stages, shows that the model can be generalized. The results found for the most generalization cases indicate R 2 equals to 0.987 and RMSE equals to 17.3 million cubic meters for training phase and R 2 equals to 0.982 and RMSE equals to 17.98 million cubic meters for testing phase. Key Words Reservoir real-time optimization, Harmony search algorithm, Fuzzy, Support Vector Machines.