Method for restoring missing data based on a combined exponential smoothing model

Abstract

Data processing and analysis are often accompanied by the problem of missing values, which can significantly affect the accuracy of predictive models and decision-making. One of the main causes of missing data in energy consumption is the periodic shutdown of systems responsible for data collection and transmission. Such disruptions can lead to data loss, complicating further processing and analysis. Therefore, an approach has been developed for filling in missing values based on exponential smoothing with adaptive coefficients determined by Fibonacci numbers. This method effectively accounts for both short-term and long-term patterns in the data, contributing to a more accurate reconstruction of lost values. This article examines a method for recovering missing data based on a combined exponential smoothing model, applied to hourly energy consumption data for the period 2016–2018. The proposed approach utilizes a regression model in which the regressors are the values of the exponential moving average, determined using smoothing coefficients and window sizes based on Fibonacci numbers. This approach effectively accounts for both new and older information by adapting weight coefficients for more accurate recovery of missing values. It has been found that when small Fibonacci numbers are used to determine the size of the sliding window and weight coefficients, the formation of exponentially smoothed values is primarily influenced by the most "recent data" (the latest, most recently obtained values). The use of Fibonacci numbers to determine smoothing parameters in the exponential smoothing method allows for the adaptive consideration of both short-term and long-term trends in time series. The proposed data reconstruction model is based on combined forecasting using six exponential smoothing models, whose parameters correspond to specific Fibonacci numbers. The combined model is built through regression analysis, enabling the adaptive evaluation of weight coefficients at each forecasting step.