Aplikace metody hlubokého učení ve vestavěných systémech

Abstract

Thesis descibes the application of deep learining with aim to train a Deep feedforward networks( also known as multilayer perceptrons), which should serve to predict solar energy availability for the following day. The prediction is based on preaviously measured atmospheric preassure values. It can be presented as regresion task, where output of this regresion is value of energy availability. Back-propagation method is used to train multilayer neural network in combination with Deep learning methods as optimalization methods(Adam, AdaGrad, Nesterov Momentum,...), Batch-Normalization method and methods for inintialization parameers of neural network. Potential application of this forecasting algorithm is in area of energy management of energy independent systems as senzors networks, especialy enviromental senzors networks. Raising efficienty of solar panels and progress in area of eletric accumulators imply raising usage of sensors networks. This systems are often applied in enviromental, where is no option, how connect this systems to primary electrical network. Solving this situation was accomplished by using battery sources to power systems. Much more interesting solution of this problem is using combination of recharging batteries or supercapacitors and source of eletric energy, which can transform some particular kind energy from enviroment to eletric energy. This aproach require ,compare to the former, another energy management, using solar availability forecasting for obtaining data, which can be used for more optimal time schedule of compute-intensive tasks and microcontroller energy modes.