This thesis develops a classification approach for the prediction of SPEs with a 48-hour lead time, and addresses the fact that very little work has been done on examining SPE forecasting methods with longer lead times than current flare-association techniques allow. Development of the technique has been based on a uniform dataset that covers 3 solar cycles and more than 30 decades of continuous spacecraft observations, and has used solar x-ray fluxes and solar ratio fluxes as predicator variables. By comparing times of SPE occurrence to times at which the solar proton flux was at a background level it has been shown that SPEs are associated with increased levels of solar x-ray flux and solar radio flux, and that these increases are, on average, significant up to 5 days prior to SPE occurrence. Using these variables as inputs neural models have generated 65% success rates for SPE prediction with a 48-hour lead time, extending the lead time of existing models by a day or more. A neural model has been coded to operate in real-time and represents the only autonomous SPE forecast model with a 48-hour lead time that does not require human supervision. Assessing the model over a 12-month operational period showed it to have superior SPE detection capability to the current 2-day forecast operated by the Space Environment Centre. Success of the classification technique was limited by the fact that solar x-ray flares were found to exhibit similar precursors to SPEs, although this meant that the model could in fact be used to forecast flares to a greater success than SPEs. Additional findings showed that the correction of radio flux observations for centre-to-limb dependence may offer the potential for more accurate forecasting ability on a timescale of days.