Several hours ahead forecast’s combination of electricity consumption based on Kuban retailer data

<p style="text-align: justify;">This thesis deals with the problem of short-term forecasting of electricity consumption. The goal of this work is to select an optimal model for hourly electricity consumption forecasting and use it for constructing day-ahead forecast.</p><p style="text-align: justify;">Hourly forecasting is very important for planning and control of power systems. It is the major instrument to determine the optimal use of generators and power plants. The commercial success of the wholesale electricity market&rsquo;s participants depends on the ability to accurately plan the necessary amount of electricity, and even a minor success in forecasting can lead to significant results on profits. Efficient trading is possible with accurate forecasting of consumption, production and imbalance between them.</p><p style="text-align: justify;">In this paper we analyzed an electricity consumption data series on an example of a Russian retailer and identified common characteristics of the series to restrict the types of the forecast models. A study was conducted on the existing methods of short-term forecasting of electricity consumption according to the data on different countries, and the class model of seasonal ARIMA was selected as the most accurate for considered data series. In this thesis were considered individual models, which include various exogenous factors and are built on intervals of different lengths; and combined models constructed as linear regressions with different weights, including the individual models and calendar indicators as factors. Based on the results of the hourly day-ahead forecasts the model with minimal error was determined. Forecasting method used in this study differs from the existing approaches in that a combination of models is constructed not only including a variety of factors, but combining models built on intervals of varying lengths; and in that a linear regression factors include indicators of the days with maximum forecast error.</p>