APPLICATION OF ADAPTIVE PREDICATIVE ANALYTICS FOR FORECASTING THE LOCAL WEATHER SITUATION

Abstract

The paper considers the main aspects of the development of a basic weather station and methods of predicative forecasting of time series in case of meteorological changes.

To build the hardware complex of the station, modern components and technologies were used: a prototype of the station was assembled based on an ESP32 microcontroller using temperature and atmospheric pressure sensor BMP180 and humidity sensor HTU2X, an architectural solution based on the LoRaWAN protocol and data exchange with the IoT TTN network was chosen. Testing the operation of the created complex confirmed the correctness of data collection and their forwarding.   

The method of forecasting local temperature changes is considered in detail using the ARIMA and SARIMA models as an example. The analysis of modern methods of forecasting time series made it possible to draw a conclusion about the feasibility of using mathematical models that take seasonality into account for long-term forecasts of temperature changes. However, these models perform worse for short-term forecasting (for the next 4-5 hours). In addition, it is shown that the use of some predicted parameters (humidity, atmospheric pressure) as exogenous leads to a deterioration of forecasting accuracy. In this case, it is worth using as exogenous well-known parameters in advance.

A hybrid method for half-hourly weather forecasting for the next two days is substantiated and programmatically implemented on the basis of standard libraries, which uses the method ARIMA (1, 2, 2) for the short-term (4-5 hours) forecast period, and then the Prophet model. The proposed solution provides quite satisfactory prediction accuracy.

Key words: time series, predicative analytics, microcontroller, data exchange protocol, temperature sensor, humidity sensor, atmospheric pressure sensor, mathematical model, time series stationarity, forecasting accuracy.