ENHANCING RESIDENTIAL ENERGY CONSUMPTION SYSTEMS USING NEURAL NETWORKS

Abstract

The article discusses current issues of energy conservation in accordance with the development program of the Republic of Kazakhstan in 2050 and the transition to a "green" economy. The paper proposes an architecture of an intelligent complex that includes photovoltaic panels, small-scale wind power generation, a battery storage system, and an inverter. For short-term hourly load forecasting, a multilayer perceptron with separate training for working, weekend, and holiday days is applied, taking into account seasonality, day type, meteorological parameters, and previous consumption data. An example of a practical implementation of an experimental installation for a residential building is presented, as well as software developed in the C# environment for monitoring, visualization, and forecasting, integrated with a DAQ data acquisition board. For profiling and classification of operating modes, an LVQ (SOM-based) network is used in combination with a system of logical decision rules that determine the power supply and energy-saving strategy. The experiment yielded the following results: the average forecasting error was 0.65%, modeling accuracy – 91%, LVQ network accuracy – 80%, sensitivity – 70%, and generalization error – 0.25. These results confirm the applicability of neural network methods as a practical tool for optimizing consumption and increasing the share of locally covered loads in the residential sector.