METHOD FOR FORECASTING LOAD AND ENERGY CONSUMPTION IN SMART BUILDING SYSTEMS BASED ON IOT DATA TIME SERIES
DOI: 10.31673/2412-4338.2025.048920
Abstract
The aim of the study is to improve the accuracy of short–term forecasting of electrical load and energy consumption in smart home systems to ensure effective proactive management of energy resources, to develop a hybrid method capable of effectively processing noisy data from IoT sensors and taking into account contextual time dependencies.
A hybrid deep learning architecture is proposed that combines convolutional neural networks (CNN) for local pattern extraction, long short–term memory networks (LSTM) for time dependency analysis, and an attention mechanism for adaptive weighting of input data importance. A feature of the method is the Context Fusion block, which integrates cyclic time features (time of day, day of the week) directly into fully connected layers of the model. Validation was carried out using the Time Series Cross–Validation method in 5 stages.
A comparative experiment on synthetic balanced data demonstrated the advantage of the proposed approach over the baseline CNN–LSTM model. In particular, a reduction of the root mean square error (RMSE) by 6.19% and the mean absolute error (MAE) by 7.26% was achieved. The mean absolute percentage error (MAPE) was less than 20%, which is an acceptable indicator for automatic control systems of household appliances. Analysis of the distribution of residuals confirmed the higher stability of the model to random emissions and noise in sensor data.
The application of hybrid neural networks for energy management tasks has been further developed by improving the mechanism for integrating heterogeneous data. For the first time, for this class of tasks, the attention mechanism is combined with separate processing of time series and contextual metadata, which allows to increase the accuracy of forecasting at peak load points without increasing the computational complexity of inference.
Keywords: Internet of Things, smart home, forecasting, hybrid neural networks, attention mechanism.
