LPWANs are networks characterized by the scarcity of their radio resources and their limited payload size. To extend the efficiency of the data transmission by decreasing the traffic sent from sensors, this paper proposes a lossy compression method using known ML techniques. We embedded a pre-trained neural network directly on constrained LoRaWAN devices and we tested the trade-off between compression ratio and accuracy of the compression algorithm. This paper studies multiple aspects of the system-energy consumption, error rate due to the lossy compression, compression ratio and the impact of LSTM parameter quantization-to measure the possible strengths and weaknesses of using a dual prediction system in order to reduce transmission costs. Surprisingly, machine learning used in this context does not consume a lot of energy and it even leads to energy saving in the very constrained devices which are the sensors.