In Software-Defined Networking, near-real-time collection of flow-level statistics provided by OpenFlow (e.g. byte count) is needed for control and management applications like traffic engineering, heavy hitters detection, attack detection, etc. The practical way to do this near-real-time collection is a periodic collection at high frequency. However, periodic polling may generate a lot of overheads expressed by the number of OpenFlow request and reply messages on the control network. To handle these overheads, adaptive techniques based on the pull model were proposed. But we can do better by detaching from the classical OpenFlow request-reply model for the particular case of periodic statistics collection. In light of this, we propose a push and prediction based adaptive collection to handle efficiently periodic OpenFlow statistics collection while maintaining good accuracy. We utilize the Ryu Controller and Mininet to implement our solution and then we carry out intensive experiments using real-world traces. The results show that our proposed approach can reduce the number of pushed messages up to 75% compared to a fixed periodic collection with a very good accuracy represented by a collection error of less than 0.5%.