A review of machine learning models and algorithms for short-term forecasting of multi-energy consumption in buildings

This article presents a comprehensive study aimed at advancing energy consumption forecasting through a meticulous systematic review of existing literature on machine learning techniques. The review synthesizes current methodologies, highlighting their strengths, limitations, and interdisciplinary applications in multi-energy forecasting, while identifying gaps that hinder precision and adaptability in dynamic environments. Addressing these deficiencies, an innovative and versatile forecasting framework is proposed and designed to accommodate the diverse energy sources prevalent in building energy supply systems, such as electrical and thermal inputs. Central to this framework is the introduction of a novel hybrid machine learning model, specifically tailored for short-term energy consumption forecasting. This model integrates advanced deep learning techniques, combining Long Short-Term Memory networks, Convolutional Neural Networks (CNN), Gradient Boosted Decision Trees (GBDT) and Transformers with a broad spectrum of data inputs, including historical consumption patterns, weather forecasts, air quality metrics and indoor conditions. By prioritizing adaptability in the training process through a dynamic, iterative pipeline, the model ensures responsiveness to evolving energy use patterns, surpassing the limitations of traditional static approaches. Results from the systematic analysis and model implementation can demonstrate significant potential of improvements in forecasting accuracy and flexibility, enabling precise predictions for both electrical energy usage and thermal comfort requirements across various building types. This research introduces a practical, scalable solution for real-world energy management challenges, contributing to the broader goals of resource efficiency and sustainable development in urban settings.