Analyzing Data Prediction in Wireless Sensor Networks

As many phenomena monitored in sensor networks are of low entropy, the prediction of such data on the receiver side has been proposed. This reduces the total number of transmissions to infrequent "model updates" and saves energy. However for evaluation, recent work unrealistically assumes reliable communication between sender and receiver nodes for model updates and also neglects the computational overhead necessary for prediction. This research shows that current evaluations of prediction schemes are quite imprecise: There is a large difference in energy consumption necessary between idealized prediction as assumed in recent work and more realistic scenarios considering packet loss rates and computational energy-costs. In addition, a simple proactive prediction scheme is proposed. Simulations indicate that this scheme is more energy-saving compared to reactive, threshold-based related work while reaching the same "Mean Relative Error".