Open-Source Machine Learning Framework for Fully Electrified Residential Load Prediction: A Temperature-to-kWh Model Using XGBoost
TL;DR
An open-source, scalable temperature-driven extreme gradient boosting (XGBoost) algorithm model is developed to predict temperature-to-kWh for residential houses and showed the novel XGBoost configuration outperformed baseline methods and yielded accuracy comparable to or better than reported deep-learning implementations for similarly sized residential datasets.
Abstract
The accelerating shift toward renewable electrification of housing, demands accurate household energy forecasting for optimizing system design, lowering costs and environmental impacts, and optimizing demand-response. To overcome the limitations of existing models, in this study, an open-source, scalable temperature-driven extreme gradient boosting (XGBoost) algorithm model is developed to predict temperature-to-kWh for residential houses. The model captures nonlinear feature interactions and works efficiently with sparse, noisy data. It is trained and validated using a five-year dataset combining hourly household electricity consumption data from a fully solar-powered and electrified residential house in Toronto Canada. The system had solar photovoltaics, photovoltaic thermal and heat pump subsystems. The results showed the novel XGBoost configuration outperformed baseline methods such as linear regression, support vector regression, and random forest and yielded accuracy comparable to or better than reported deep-learning implementations (e.g., multivariate long-short term memory) for similarly sized residential datasets. The results for the model trained with 5-year datasets show RMSE of 0.13 and MAE of 0.06 with the spike handling process, which reflects the model’s reliability for fine-grained residential energy forecasting. This new model can be used to improve optimization and modelling tools to enable a wider deployment of solar-powered fully-electrified residences.