Seite - 132 - in Intelligent Environments 2019 - Workshop Proceedings of the 15th International Conference on Intelligent Environments

estimation of yields, identification ofweeds, predictions on climatology, etc. For each of thesefieldsof application, different typesofdeep learning techniques areused, high- lightingconvolutionalnetworks for imageclassificationandLSTMfor thepredictionof temporalvariables. In literature we can find some works that have tried to predict the climatic vari- ablesusingdifferent climaticvariablesobtainingpositive results.Thus, in [16], thecon- structionofa robust statisticalmodel isproposedforpredictingmeteorologicalvisibility based on other intermediate variables (temperature, pressure, humidity and dewpoint). Two single-layer and four-layer LSTMnetworks are used. The data have been prepro- cessedbymeansofnormalization, rescaling to the range[0,1]andusingamovingaver- age. ThemultilayeredLSTMmodel proves to be themost effective.Anotherwork that uses anLSTM to predict climate variables is presented in [19]. The variables used are temperature, humidity andwind speed. In this case the network architecture consists of twoLSTMlayers.Theactivation functionchosen for theoutputof thedense layer is the RELU.Theoptimizerused isRMSProp.Thedatahavebeennormalizedand rescaled to the range [-1, 1] and the results obtainedhavebeen satisfactory. In [11], the authors in- tend tomodel rainandrunoffusingLSTMnetwork,whichpredictdischargeforavariety ofwatersheds.The authors aim todemonstrate thepotential of thismethod. Somevari- ablesusedareday length, rainfall, temperatureorhumidity.Thenetwork iscomposedof 2LSTMlayersandbetween themaDropout layer toavoidovertrainingof thenetwork. Thedifferencebetween the existing techniques and theproposalmade in thiswork is that here we try to predict the temperature value in order to be able to incorporate an intelligent component into an IoT systemandalso that only the temperature value is used. 3. Materialandmethod 3.1. DataCollecting The data used to train and test the proposedmodel are real data obtained from the IoT systemdeployed in an agricultural plot ofMoratalla, a village in theRegionofMurcia, Spain. The deployed IoT system consists of several nodes located in different areas of theplotwith sensorsof temperature, humidity andwind speed.The IoTnodesuseLora technology to communicate and send the data viaGPRS to a data visualization appli- cation. In addition these data are preprocessed to avoid the errors that can be proved by the sensors.The errors havebeen eliminatedusing theKmeans algorithmfor outlier detectionandcorrectionpresented in [5]. Thedatausedcorrespond to theperiod from1/11/2018 to28/02/2019, having tem- perature values (measured in degreesCelsius [◦C])with a frequency of 10minutes. A totalof17212 temperaturemeasurementsareavailable,where foreachvalue it isknown thedateand timeof themeasurementand the temperaturevalueat that time. 3.2. DeepLearningLSTM LSTMs are a model of Recurrent Neural Networks (RNN) series capable of learning long-termdependencies[8,9].This isaproblemthatyoupresent theRNN,sinceitworks M.Á.Guillén-Navarroetal. /AnLSTMDeepLearningScheme forPredictionofLowTemperatures132