A distributed data compression algorithm for low-power wide-aria networks is proposed. The algorithm is based on prediction of observed process on server side with controlling prediction error on the end device side. The prediction is completed by recursive linear prediction algorithm using Levinson-Durbin recursion. The algorithm provides obtaining the estimated values in real time on server side with no sending any data through the network until prediction error exceeded threshold. It allows to use wireless transceiver less intensively saving the battery budget this way, and to reduce number of packets and its length saving the wireless channel capacity. The efficiency of algorithm is investigated on end device and server models using the real data collected by CO2, humidity and light sensors.