DEVELOPMENT OF A WEATHER FORECASTING SYSTEM USING THE LSTM MODEL FOR SMART HOME SYSTEMS

Abstract

Due to rapidly increasing demands for energy efficiency and the growing level of automation in household and industrial systems, highly accurate meteorological forecasting is becoming particularly important. Traditional statistical methods are often insufficient for modeling local weather processes, underscoring the need to employ recurrent neural networks and other deep learning approaches capable of capturing complex temporal dependencies in the data.

The study conducted in the paper includes an analysis of scientific sources related to the research topic, which indicates that deep learning methods have been actively employed in recent years for various forecasting tasks, including weather prediction, and this area continues to attract considerable attention from researchers. To implement the LSTM model, the mathematical foundations of the core architecture of this recurrent neural network were examined, incorporating an attention mechanism to enhance forecasting accuracy by enabling the model to focus on the most informative segments of the input time series. The model proposed in this paper is multitasking, as it produces six independent meteorological outputs – five regression variables (temperature, pressure, humidity, wind speed, and cloudiness) and one classification variable. In this study, the precipitation probability is implemented as a binary classification feature indicating the presence or absence of precipitation, which enables a comprehensive assessment of the multitask architecture. Algorithms for meteorological data preprocessing and model training are proposed, while forecast accuracy is evaluated using the standard MAE and RMSE metrics. The authors developed software based on the Python programming language, the Flask framework, and the TensorFlow and Keras libraries to provide convenient user interaction with the proposed LSTM model.

The conducted experimental studies and comparative analysis of the obtained results demonstrated that the LSTM architecture with an attention mechanism achieved the best error metrics compared to the forecasting results produced by the Random Forest and Ridge Regression models. This confirms that the advantages of the LSTM model lie not only in its nonlinear modeling capability but also in its inherent temporal memory, which is critical for accurate modeling of weather processes.