Towards net zero energy building: AI-based framework for power consumption and generation prediction

Abstract

Net zero energy buildings (NZEBs) are at the forefront of sustainable construction, aiming to balance energy consumption with energy production over the course of a year. These buildings minimize environmental impact by integrating advanced energy-efficient technologies and on-site renewable energy sources. However, challenges, including mismatches between energy demand and supply, influenced by consumer behavior and weather conditions, can disrupt the effective management of smart grids. To meet this challenge, an efficient AI-based approach is highly desirable to assist the policy maker and power management. Therefore, we present a two-stream network for short-term consumption and generation prediction, composed of three main tiers: primarily, a preprocessing technique is employed to refine the raw data collected from smart meter and fronius smart meter. Next, these polish records are passed individually to two independent networks, such as the energy convolutional network (ECN) and the temporalflow network (TFN), responsible for extracting spatio-temporal energy up-down patterns simultaneously. Later, the resultant features are aligned using a fusion mechanism for the final outcomes. Numerous hybrid models performance are analyzed to fairly evaluate the strength of the proposed model. The experimental study indicated that our model achieved a notable reduction in the mean squared error (MSE) of 0.0075, 0.035 and 0.030 on hourly data from the data sets of building consumption (BC), photovoltaic (PV) and intermittent renewable daily electricity (IRDE), respectively, exceeding the performance of current state-of-the-art methods (SOTA). Existing AI-based techniques typically utilize dedicated networks tailored for specific tasks, focusing either on consumption or generation prediction. In these models, CNNs extract spatial features directly from real data and subsequently learn temporal patterns. In contrast, we proposed novel network and learning mechanism that simultaneously extract spatio-temporal features for both consumption and generation prediction. This generalize approach enhances the model's efficiency for power matching, offering greater versatility and effectiveness compared to existing solutions. • A two-stream parallel learning framework is proposed for net zero building power matching. • An ECN is developed to extract spatial features from solar and building data. • A TFN is specifically designed to effectively extract energy temporal information. • Comprehensive experiments are conducted using PV, BC, and IRDEP datasets.

Topics & Concepts