This paper describes a feature-based pattern-recognition technique that utilizes real-time solar wind measurements to identify and predict the occurrence of solar wind structures that can cause geomagnetic storms. The technique is based on the knowledge that storms are caused by solar wind events with certain identifiable features, the two most important ones being (1) extended periods and (2) large amplitudes of southward interplanetary magnetic field (Bz<0). Using the measured solar wind properties available at the current time t the profiles of the north-south magnetic field component, Bz(t⁁), and the east-west component, By(t⁁), are estimated for the solar wind that has yet to arrive, where t⁁ is the future time. On the basis of the estimated Bz and By profiles the occurrence, onset time, duration, and severity of impending storms are predicted. It is shown that the occurrence and durations of large storms exceeding a predetermined threshold can be predicted accurately. Successful predictions can be made after examining the initial ~20% of a geoeffective (i.e., storm-causing) structure. For long-duration southward Bz events such as magnetic clouds the advance prediction time can be from several hours to in excess of 10 hours. The technique is tested using 5 months of historical solar wind data. We develop a procedure to extend the capability of the technique by incorporating additional features. Reduction of potential false alarms and misses resulting from such extensions is discussed.