Prediction Models

Chapter 7 deals with the approximation of internet load prediction using Julia. First, we explain the current state of model development. Classical prediction models with linear approximations are being replaced by nonlinear models with neural networks. After describing the essential model features, representative examples are presented. The perceptron network as a basic element of approximation is implemented. We explain the learning formulas and immediately convert them into Julia program code. After explaining the learning algorithm, we save ourselves further programming of more complex learning tasks, as this is taken over by the comfortable Julia module Flux.jl. The potentials and limitations of the forecast are analyzed using uncorrelated and correlated artificially generated as well as real data. For independent work with the programs, the removal of outliers using the Tukey method is suggested before the prediction calculation. A real data set with load measurements in minute intervals over 66 days is provided for pattern search.