In this paper, a mode transformation algorithm based on traffic prediction in virtual multiple optical line terminal (OLT) passive optical network (PON) is proposed. By proposing exponential smoothing algorithm based on weight update (WU‐ESA), user traffic is predicted well. WU‐ESA is a combination of two algorithms: exponential smoothing algorithm (ESA) and genetic algorithm (GA). The weight in ESA is optimized by GA based on real‐number encoding. By setting two periods, GA part and ESA part can be separated effectively. By presenting elastic packing algorithm (EPA), the load balance problem in virtual multi‐OLT PON is solved. EPA is implemented based on WU‐ESA. By the simulation and analysis, the effectiveness of the proposed algorithms is demonstrated. Compared with traditional mode transformation algorithm, EPA shows good performances in delay, throughput, and packet loss. Compared with traditional mode transformation algorithm, the EPA makes the packet loss decrease by at least 5% when the system load is greater than 0.9. Meanwhile, the delay of the two subsystems can be kept at a relatively balanced level by the implement of EPA. For throughput, the use of EPA improves the throughput by 20% when the system load is high.