In Perm, a new method has been discovered to enhance the efficiency of the city's heat supply system.

As the country prepares for winter, the issue of heating supply in apartment buildings becomes increasingly relevant. Delivering heat transfer fluid to consumers requires maintaining the temperature at a specific level, taking into account the weather forecast for the upcoming period. To ensure optimal operation of the heating equipment in boiler houses, intelligent management systems based on machine learning methods are employed. These systems require regular training to adapt to changes in the technical condition of the heating network. Researchers from Perm Polytechnic University, in collaboration with colleagues from SoftM, have developed and tested an intelligent module that incorporates the most effective models for adjusting the results of neural network forecasting. This advancement aims to reduce the risk of errors and enhance the efficiency of energy resource consumption.

2024-10-15 07:32:27https://naked-science.ru/article/column/sposob-povysit-effektivno

The article was published in the journal "Bulletin of South Ural State University. Computer Technologies, Management, Radio Engineering." The research was conducted as part of the strategic academic leadership program "Priority 2030."

The heating supply of residential premises with hot water is carried out through the pipes of the central heating system. The boiler room serves as the heat source where water is heated and then supplied to the heating node of the district. To optimally regulate the operation mode of the boiler room, the heating organization can employ various methods and management systems. For instance, for gas boilers, automated systems are implemented to maintain the set output temperature by regulating the boiler's operation and fuel supply according to the required parameters. An efficient fuel combustion mode allows for a reduction in energy carrier (gas) costs and enhances the economic and ecological effectiveness of the process.

During the operation and maintenance of the heating network, its properties change, leading to increased or decreased heat losses that affect the accuracy of the control model. To compensate for these changes, it is necessary to periodically retrain the neural network model so that it can predict the network's performance considering the ambient temperature and the technical condition of the heating network. However, this requires significant time investment.

Therefore, scientists from Perm Polytechnic University have for the first time utilized and compared two models for refining the results of predictive neural network control, as well as analyzed the effectiveness of each. A statistical regression linear model was considered due to its high accuracy and simplicity in training, alongside the decision tree-based model XGBoost. The latter is a graphical scheme consisting of nodes, leaf nodes, and branches that describe the probabilities of event developments. Each subsequent branch is designed to correct the error of the previous one, reducing the mean deviation. This process continues until the error is minimized or one of the early stopping rules is met.

For training and testing the models, the polytechnic researchers selected 10 apartment buildings, for which the data over a certain period had the least number of gaps due to technical reasons. A separate model was built for each building, which was then used to calculate the temperatures of the heat carrier at the entrance to the apartment buildings. The results were subsequently compared with the actual values from the specified sample.

“The maximum deviation of the calculated temperature from the measured one in XGBoost was 4.8 °C, while in the linear model it was 6.1 °C. This means that the former is more effective, as its error magnitude is significantly lower. The proposed methods have been tested on real data, confirming their applicability in the development of an intelligent information system for heating management,” comments Valery Stolbov, professor of the Department of "Computational Mathematics, Mechanics, and Biomechanics" at PNIPU, Doctor of Technical Sciences.

PNIPU scientists have identified the most effective model for predicting the behavior of the heating network, which will enable the correct selection of control actions. This will significantly reduce the risk of violating environmental regulations and the waste of resources on excessive fuel and electricity consumption, as well as on equipment maintenance and repair.