Increasing accessibility to advanced sensors and meters, wide implementation of variable speed drives and intensive development of optimization methods open new ways to control energy conversion processes. Complex pumping stations working at wide ranges of flow conditions can be controlled various ways to meet requirements. In recent years researchers focused their interests on energy efficiency of such systems. The main goal of this research is to establish fundamental knowledge leading towards smart, self-learning control applications of multi-pump systems in parallel configuration. This goal is achieved by finding an answer for the following research questions (RQ). RQ1. Which control methods are the most energy-efficient? RQ2. What regularities in system operation could be the objective of self-learning systems? RQ3. Which measurements and control equipment are necessary for effective self-learning control methods? Addressing these questions, a set of optimization tasks by genetic algorithm (GENOCOP) have been performed for a pumping station with four parallel connected pumps having various characteristics. Each pump can be individually controlled by a combination of three methods: (i) variable speed drive, (ii) discharge valve, and (iii) by-pass flow. A real experimental pumping station provided input characteristics as well as reference measurements for verification of obtained numerical results. Both numerical simulations and verifying experiments have been performed for a combination of a wide range of flow rate requirements (10-90% of pumping station nominal flow) and five various load resistance characteristics. Completed work demonstrated that the most energy efficient control of a complex pumping station can be achieved by a set of variable speed drives powering pumps individually. Clear regularities have been shown across a range of flow rate requirements and load resistances. The most energy efficient operations can be obtained by pumps working at their maximum count and lowest possible frequencies. In the case of pumps that have identical characteristics, all variable speed drives should operate at the same level. However, for pumps with various characteristics, more energy efficient operation can be achieved by additional tuning of individual frequencies. With current computing capacity, the methodology used in presented research is impractical to integrate into a real-time control loop. Therefore individual tuning can be an object for further research towards determining hardware and software structure supporting self-learning algorithms to control complex pumping stations.