Investigating Design Parameters for Practical Load Forecasting of Grid-Interactive Buildings Using LSTM

Abstract

The dissertation presented contributes to research in residential load forecasting using machine-learning (ML) models to optimise energy management within grid- interactive efficient buildings (GEB) and addresses the challenge of implementing forecasting models for practical applications. The use of accurate load forecasting ML models has been shown to extend and im- prove the field of energy management towards the resource coordination of GEB. Developing these models is a time-, data- and compute-intensive process that re- quires selection and tuning of various design parameters. In practice however this has costs and limitations: collection and storage of real-time data, measurement device costs and finite compute power. To simplify model development and reduce these costs, this dissertation investigates to what extent the relationships between model parameters and performance can inform ML design decisions. In the research presented, 3 experiments are performed to investigate the relation- ships between training data size, prediction horizon length and data resolution and their impact on the predictive model fidelity. Several prediction models are devel- oped using various parameter settings. The accuracy and compute intensity are measured to evaluate the influence of each parameter on the model performance. A dataset of 3 years and 11 months of residential power consumption measure- ments at a one-minute resolution is used as the load profile for this investigation. For this load profile it is found that increasing training data size increases compute time linearly, with an exponential decay in model prediction error. This results in a maximum, resource-efficient training data size of 450 days. In addition, the effec- tive prediction (i.e prediction models with R2 scores greater than 0) horizon length is found to be 3 hours. At this length the load profile is suited for short-term load forecast applications such as distributed energy resource (DER) and real-time pric- ing (RTP) management. Furthermore, models trained on low-resolution data (up to 30 minutes) can achieve comparable performance to higher-resolution models with at least 3 additional months of training data. The findings of this investiga- tion therefore represent a contribution towards the development of a ML design tool for more efficient design parameter tuning considering practical load forecast- ing conditions. This shows potential to enable more efficient implementations of load forecasting systems for GEB environments.