An Energy Consumption Forecasting Tool for Buildings Based on Multivariate Deep Neural Network Model

Buildings could play a crucial role in the coordinated management of smart energy grids to achieve sustainability objectives. However, to effectively reduce end-use energy demand and enhance buildings' readiness level and flexibility for grid integration, there is a need for technological enablers that facilitate cost-effective digital upgrades of buildings and energy prediction mechanisms. In this paper, we present an energy prediction tool that utilizes a multivariate deep neural network model to forecast the day-ahead energy consumption of buildings. This tool can interface with physical IoT monitoring devices in buildings via message queues or brokers, as well as with existing datasets. It is designed to handle high volumes of raw data and diverse data sources using a MongoDB database. The prediction model considers both historical data and newly sampled data comprising multiple layers of interconnected neurons arranged in a multilayer perceptron architecture. Each layer applies a series of weighted transformations to the energy and weather input data, enabling the model to learn patterns and relationships for accurate energy predictions. The developed tool and model successfully address state-of-the-art challenges regarding the integration of heterogeneous IoT energy data sources and the predictive power of machine learning showing good accuracy results for different buildings.