Abstract
<jats:p>The existing approach of “top-down” energy consumption forecasting does not take into account the actual state of engineering systems and enclosing structures, which causes unjustified losses and inefficient allocation of energy resources. The relevance of the study is the need to bridge the gap between the design documentation and the actual operation modes of the housing stock, which leads to systemic errors in urban energy planning and excessive redundancy of capacity at generation sources. The aim of the study is to improve the planning of the city energy consumption, minimize the error of forecasting daily and seasonal heat load schedules based on verification of the actual parameters of energy consumption of buildings and the transition to the concept of “bottom-up” modeling. The paper uses methods of mathematical statistics (correlation analysis), analytical method, as well as methods of linear regression (LR), random Forest and neural networks of long short-term memory (LSTM). A study of energy consumption based on 15 residential apartment buildings has been conducted. It has been revealed that the actual energy consumption exceeded the design consumption by an average of 16,2 %. It has been found that the real energy efficiency class of buildings is lower than stated in the certificate in 86 % of cases. Based on the data of the design documentation and the actual readings of heat metering units, a hierarchical model to forecast energy consumption has been developed. The implementation of the developed methodology to forecast energy consumption of buildings and LSTM-based models will reduce the average absolute error by 7,5 times compared to standard calculations. It is proved that verification of actual consumer parameters is a critical factor to improve forecast accuracy, providing a basis for optimizing urban heating networks and generation capacities, which will ensure improved planning of the city energy consumption.</jats:p>