Determination of Hyperparameters in the Development of a Frost Predictive Model with Data Science

The damages caused by the meteorological phenomenon of frost on crops result in significant economic losses. Farmers consistently seek assistance from experts or technological tools to forecast the occurrence of the phenomenon and protect crops. Following the Data Science process, the values of climate variables recorded by two weather stations have been preprocessed. Various techniques were necessary to generate datasets suitable for the developed predictive models. Sliding window was used for the data. The window size required a particular experiment. The models are based on an LSTM neural network. The learning rate for the network was obtained from a set of tests, in which metrics and graphical representations were analyzed. The models predict the temperature for a horizon of three hours. The best model operates with a three-hour sliding window, achieves a precision of 92% and 73% for recall for frost cases.