Real-time production monitoring in the oil and gas industry has become very significant particularly as field operations become economically marginal with increasing reservoir depletion. Production measurements are typically performed with conventional test-separator facilities, which don’t deliver continuous production information. An alternative solution are physical multiphase flow meter. Their application is desired but costly and requires a good understanding of the governing system physics and fluid chemistry. This work describes the development of an innovative metering technology, which is known as Virtual flow meter (VFM). VFMs are data-based mathematical models for real-time multiphase flow prediction, which make use of readily accessible sensor readings from wells. For this work different physical properties were measured and recorded in a controlled laboratory environment to provide input data for learning, validation and testing of an artificial neural network (ANN). The experiments were conducted in the Pump testing-facility (PTF) using a vertically installed electric submersible pump (ESP) at the Montanuniversity in Leoben. The objective was to construct a three-phase flow loop were different flow configurations could be tested and quantified under firm, reliable and repeatable laboratory conditions. The idea was to analyze a broad spectrum of different flow conditions, by manipulating the flow rates and simultaneously record the sensor responses installed along the flow path. Both, sensor data and flow measurements were then processed and used as input for the mathematical model. The experimental program consisted of 32 experiments, 3 single-phase, 11 two-phase and 18 three-phase experiments. In total 85 different flow configurations could be investigated. All three-phase experiments consisting of 19 different records and flow rates of water, synthetic oil and pressurized air were implemented in the VFM model for multiphase flow prediction. Initially, every single experiment was investigated and modeled individually to analyze the differences in flow prediction accuracy as a function of the flow rate of each phase. Finally, the processed data of all 18 three-phase experiments were modeled together in three separate neural networks with water, oil and gas as outputs, to avoid any interference between the predicted flow rates due to different output ranges. The reached prediction accuracy of the phases is technically useful and results in an average relative error of 1,20%, 4,85% and 2,40% for water, oil and gas respectively. The measured flow rate ranges are between 0-12 m³/h for water, 0-2,8 m³/h for oil and 0-18 kg/h for gas. The created model can predict flow rates at reasonable flow rates and proves the potential of sensor-data in multiphase flow prediction and is herewith capable to monitor production outputs in real-time.