Implementation of a neural network model in the Statistica 12 for mudflow frequency forecasting

The article describes some principles of operation of an artificial neural network. It provides an example of implementing a neural network model by selecting its best architecture using the Statistica 12 software package. The article considers a method for neural network forecasting of a series of mudflow events based on nonlinear relationships with precipitation and temperature series. To solve the problem, the Data Mining (intelligent data analysis) block – Neural Networks was used in the Statistica 12 package. A multilayer perceptron (MLP) was chosen as a neural network method, and a hyperbolic tangent (tanh) was used as an activation function. Based on deep learning algorithms, a mathematical model MPL 2-50-1 was developed, which is capable of learning on the used data (precipitation, temperature, number of mudflows for the period 1953-2015) and forecasting the number of mudflows based on the meteorological parameters (precipitation, temperature) entered into the model. It was found that with average precipitation values of more than 110 mm in the period from May to September from 2016 to 2034, the number of mudflows is predicted to be from 10 to 13, which is higher than their average value of n = 8 for the period with actual initial data from 1953 to 2015. Trends in the number of mudflows in the Terskol Gorge in the warm season from 1953 to 2015 (the period with actual data) and from 2016 to 2034 (the period with predicted data) were determined using polynomial and linear trends. It follows from the linear trend equation that, on average, over the entire period, including the predicted one, the number of mudflows tends to grow slightly by 0.3/10 years. The polynomial trend demonstrates an increase and decrease in the number of mudflows at different time intervals. In the forecast interval of 2016-2034, the decrease in the number of mudflows demonstrates both a polynomial trend and a linear trend.