Solving Some Problems of Predictive Analytics for Time Series Data

Approaches for constructing an additive regression model and a seasonally integrated autoregressive-moving average model were considered. Experiments were aimed at investigating and developing an algorithm for forecasting meteorological fields. In particular, a time series forecast of monthly precipitation totals for the next two years was built. In the software experiments for building forecast models, arrays containing data from eight-year observations of atmospheric precipitation were used. That is, sums of precipitation for the period between standard synoptic dates with an interval of three hours were used. The year 2016 was selected as the forecast year. The original data series was processed - emissions that were not attributed to seasonal variations were removed. Additionally, the allowable maximum amount of precipitation between dates for an observation station with synoptic index 29430 recommended by the All-Russian Research Institute for Hydrometeorological Information - World Data Centre (VNIIGMI-WDC) was controlled. The Python programming language was used as a tool to generate forecasts for the time series data. An additive regression model was created using the Prophet library from Facebook. A seasonally integrated autoregressive-moving average model was created using the StatsModels library. A DF-test (DickeyFuller basic test), which checks for the presence of a single "unit root", was used to test for stationarity. The testing and writing of the program code took place in the interactive Jupyter Notebook environment. The environment is a graphical web shell for Python and extends the idea of the console approach to interactive computing. Model forecast accuracy was assessed by calculating absolute and mean absolute prediction errors.