In dissertation, the main attention is paid to the development and improvement of the prototype of the hardware and software complex for the analysis and processing of data of an intelligent house. The analysis of scientific publications and open sources on the topic of intelligent buildings showed that there are currently not enough solutions for hardware and software systems, the functioning of which does not depend on cloud technologies and access to the global Internet. A separate and no less important component in the design of intelligent building control systems is the consideration of thermophysical processes, which are described by the corresponding dynamic equations of heat transfer in the room. Solving such equations allows both modeling of thermal processes and the use of their solutions to refine forecasting of energy consumption. The first chapter of the dissertation examines the problems of creating intelligent home systems, which are mostly based on boundary and fuzzy calculations. The current state and aspects of edge computing for the intelligent home, based on Raspberry Pi 3 microcomputers and Arduino microcontrollers, are analyzed. Also, current means and approaches to sensor data processing within regression and neural network models are considered, the latter of which can be built into microcontrollers for the purpose of prediction. To ensure flexible deployment of neural network models, a review and analysis of automated data processing cycle tools — training and validation of a neural network — embedding a neural network on a microcontroller with simultaneous deployment on a microcomputer was conducted. The second chapter of the dissertation describes the data used in system modeling. Algorithms for the detection of emissions and anomalies were considered and implemented for the processing of these data; statistical analysis was carried out and regression models were built: a) external and internal temperatures; b) gas consumption and battery surface temperatures and c) external, internal and heating element temperatures, the results of which can be used for forecasting and analysis of relationships between measurements. An approach to the disaggregation of gas consumption data based on cluster analysis is proposed, the purpose of which is to select specific gas consumers from aggregated (aggregated) data. For the construction of neural network prognostic models of temperatures, the process of data preparation and transformation is described in detail, and current architectures of recurrent neural networks, on the basis of which the corresponding models are implemented, are considered. The problem of hyperparameter optimization of designed neural networks is formulated, the solution of which is carried out using a genetic algorithm. Analysis of the results and selection of the optimal architecture for deployment on the STM32 microcontroller was carried out. The third section of the dissertation is devoted to the hardware and software implementation of the complex prototype for analyzing and processing data of an intelligent building. The Nvidia Jetson Nano microcomputer, which has sufficient power for a central computing hub, and a high-performance STM32 F767 microcontroller were chosen as the hardware platform. The latter is the basis of primary boundary calculations due to the neural network deployed on it. When temperature sensors are connected to it, the neurocontroller can make short-term predictions in real time for comparison with the data coming to it (in case of anomaly detection), or be used as a separate predictive module to optimize room heating parameters. The software part of this prototype consists of a REST API wrapping a database with SQLite measurements, modules for statistical processing and neural network modeling, a system for organizing the MLOPS pipeline — Mlflow, an Apache Airlow orchestrator, and a communication module with a neurocontroller based on TCP sockets. In the fourth chapter, a method of applying thermophysical models in combination with temperature and gas consumption data is proposed for an approximate estimation of the effective coefficients of thermal conductivity and thermal capacity of the building. Calculation of such parameters takes place on the basis of solutions of direct and inverse problems by formulating the problem of optimization of the functional, which determines the difference between the calculated and real temperatures in the room relative to thermophysical coefficients. Also, a mathematical model for evaluating the effective thermal parameters of individual heating sources has been developed, which can be used both for consumption forecasting tasks and for determining the part of the total heating energy consumed by a specific heating element.
