Previous work has pointed to the physical mechanisms behind the nocturnal offshore propagation of convection south-west of Sumatra. Low-level moisture flux convergence due to the land breeze front controls the progression of a squall line away from the coast overnight. However, the diurnal convection over the mountains occurs on only 57% of days in December-February (DJF) and propagates offshore on only 49% of those days. We investigate day-to-day variability in dynamical and thermodynamical conditions to explain the variability in diurnal convection and offshore propagation, using a convection-permitting simulation run for 900 DJF days. A convolutional neural network is used to identify regimes of diurnal cycle and offshore propagation behaviour. The diurnal cycle and offshore propagation are most likely to occur ahead of an active Madden-Julian Oscillation, or during El Niño or positive Indian Ocean Dipole; however, any regime can occur in any phase of these large-scale drivers, since the major control arises from the local scale. When the diurnal cycle of convection occurs, low-level wind is generally onshore, providing convergence over the mountains; and low-level humidity over the mountains is high enough to make the air column unstable for moist convection. When this convection propagates offshore, mid-level offshore winds provide a steering flow, combined with stronger convergence offshore due to the land breeze or convection-triggered cold pools. Low-level moisture around the coast also means that, as the convection propagates, the storm-relative inflow of air into the system adds greater instability than would be the case on other days.