IN-24-C025 - Miscellaneous Electric Loads Prediction By Deep Learning Implementation: An Educational Case Study PDF

Apart from a large energy consumption due to the Heating, Ventilation, and Air Conditioning (HVAC) loads in buildings, the contribution of standard equipment in buildings is approximately 17% of the load consumption. Understanding consumption patterns would be important for achieving energy savings in addition to conducting accurate predictions for Building Management System (BMS) and Model Predictive Control (MPC). However, Occupant-centric Miscellaneous Electric Loads (MEL) are rarely explored, specifically considering the exposure to extreme climatic conditions and specific building types (such as educational). In addition, they are hard to predict; in particular, considering the condition that occupants are in control of the devices in shared spaces. Implementing three distinct deep learning models, including Long Short-Term Memory (LSTM), a deep neural network model, and Bayesian LSTM, this research evaluates and compares their performance in accurately forecasting MEL in an educational case study in the UAE. Notably, Bayesian LSTM demonstrates the highest level of predictive accuracy with minimal deviation from measured values among the models considered. However, the computational expensiveness of LSTM and Bayesian LSTM compared to DL and trade-offs of performance and time should be an obvious consideration for real implementations. In practical terms, these predictive models illuminate the sources of energy usage, emphasizing the importance of comprehending various building plug load devices. Deep learning models, particularly when applied as time-series data analysis techniques, provide a robust means to answer how can we make accurate predictions on electric plug loads, facilitating the analysis and reduction of MEL consumption. These predictive models offer actionable insights for BMS and empower decision-makers to implement intelligent energy management strategies.