Predicting IoT Readings With Deep Learning

Rad se bavi problemom predviđanja očitanja uređaja interneta stvari. Objašnjeni su koncepti interneta stvari, dubokog učenja i povratnih neuronskih mreža. Napravljene su tri različite arhitekture modela povratnih neuronskih mreža koje se razlikuju po tipovima ćelija. Korištene su osnovna RNN ćelija, LSTM ćelija i GRU ćelija. Proveden je postupak pretraživanja po rešetci kako bi se pronašli optimalni hiperparametri. Postupak je proveden nad jednostavnijem datasetu koji sadrži samo značajku temperature zraka i na potpunom datasetu koji sadrži sve dostupne vremenske značajke poput temperature, tlaka i vlage zraka, brzine i smjera vjetra i slično. Predstavljeni su i analizirani rezultati istreniranih modela te je razvijen sustav za alarmiranje na ThingsBoard IoT platformi koji aktivira alarme pri predviđanju temperature koja izlazi iz definiranih raspona.The paper deals with the problem of predicting Internet of Things device readings. Concepts of Internet of Things, Deep Learning and Recurrent Neural Networks are explained. Three different architectures of recurrent neural network models were created, differing by cell type. Basic RNN cell, LSTM cell and GRU cell were used. A grid search procedure was performed to find the optimal hyperparameters. The procedure was performed on a simpler dataset containing only the air temperature feature and on a complete dataset containing all available weather features such as air temperature, air pressure, air humidity, wind speed and direction, and others. The results of the trained models were presented and analyzed, and an alarming system was developed on the ThingsBoard IoT platform that activates alarms when predicting a temperature that goes out of the defined ranges