ADAPTIVE MODELS OF ELECTRIC LOAD FORECASTING OF PRODUCTION CONSUMERS

Abstract

The article investigates models and methods of electric load forecasting. It is shown that the following methods of power consumption control are currently known: instantaneous norm; at the ideal rate; management on the forecast value; control with the use of average power on a moving time interval ("moving window" method). It is shown that it is better to focus on those methods that are based on the study of forecast estimates, which are the source information for management decisions. The main requirements for real-time systems are: high accuracy of operational forecasting and simplicity of algorithms, which provides a minimum solution time; work in conditions of uncertain and insufficient information, ensuring the stability of management. The analysis of works devoted to the issues of forecasting the processes of power consumption management systems of industrial enterprises is carried out. It is shown that automated control systems have specific requirements for mathematical forecasting methods due to little study of the nature of the forecast parameter, small amount of reporting statistics and insufficient reliability of source information and most accurately meets such requirements, adaptive approach to method design. The adaptive approach allows to solve the problem of adequacy of the method of the object of forecasting and from the point of view of simplicity of realization and time of calculations in the first place it is necessary to put adaptive methods of forecasting and, first of all, the method of exponential smoothing. Exponential smoothing, considered in this paper as a predictive model, allows to identify the inadequacy of the model to the real process and to bring the estimate of the determined basis of the process closer to the real one, ie to reduce the prediction error. However, this requires time, which increases with increasing changes in the coefficients of the model. In this regard, there is a problem of regulating the reaction rate of the predicted model to changes in its coefficients. A number of methods of automatic adjustment of the smoothing parameter are considered and analyzed: the evolutionary method of adaptation, methods using the tracking signal, methods of adapting the parameter by, optimization using gradient smoothing. It is shown that the method of adaptation using the tracking signal is simple and especially valuable for modeling series with a short history. Due to its simplicity, this method is especially convenient where predictions are made using computer technology. Studies of this model of forecasting on statistical data obtained at various enterprises have shown that the model adapts to real data at step 4-6 of forecasting and then the forecast error does not exceed 2%. Analysis of adaptive forecasting models based on the method of exponential smoothing showed their high efficiency and good adaptability to changes in the process of electricity consumption. The greatest difficulty in forecasting are cases of abrupt changes in the development of the process. Abrupt changes in the process can lead to a violation of the previously existing qualitative relationships of the parameters of the projected system. If there is a jump, it is very important to assess whether the deviation is caused by an obstacle or whether it is due to a change in the predicted process. If the changes are caused by an obstacle, it must be filtered out. If the deviations are caused by a change in the model, then the current process data are of the greatest value. From the point of view of fast working off of abrupt change the model of exponential smoothing with high value of the smoothing parameter is rather effective. However, this model is highly susceptible to interference. To eliminate this circumstance, a modified procedure for correcting the parameters of the forecasting model is proposed. The procedure is based on the introduction of a logical operator, which is based on the analysis of inconsistencies in the forecasts and imposes additional restrictions on changes in the smoothing parameter and the values of the original statistics. Experimental studies of the considered models are carried out.