We study the modelling of large data sets of high frequency returns using a long memory stochastic volatility (LMSV) model. Issues pertaining to estimation and forecasting of datasets using the LMSV model are studied in detail. Furthermore, a new method of de-seasonalising the volatility in high frequency data is proposed, that allows for slowly varying seasonality. Using both simulated as well as real data, we compare the forecasting performance of the LMSV model for forecasting realised volatility to that of a linear long memory model fit to the log realised volatility. The performance of the new seasonal adjustment is also compared to a recently proposed procedure using real data.