Wireless sensor networks (WSNs) are deployed to monitor physical events such as fire, or the state of physical objects such as bridges in order to support appropriate reaction to avoid potential damages. However, many situations require immediate attention or long-reaction plan. Therefore, the classical approach of just detecting the physical events may not suffice in many cases. We present a generic WSN level event prediction framework to forecast the physical events, such as network partitioning, well in advance to support proactive self-actions. The framework collects the state of a specified attribute on the sink using an efficient spatio-temporal compression technique. The future state of the targeted attributes is then predicted using time series modelling. We propose a generic event prediction algorithm, which is adaptable to multiple application domains. Using simulations we show our framework's enhanced ability to accurately predict the network partitioning with very high accuracy and efficiency. Copyright © 2012 Inderscience Enterprises Ltd.