Data-driven approach to prediction of residential energy consumption at urban scales in London.

Development of energy prediction models plays an integral part in management and enhancement of the energy efficiency of buildings, including carbon emission reduction. Simplified and data-driven models are often the preferred option when detailed information of simulation is not available and the fast responses are required. This study developed data-driven models for predicting electricity and gas consumption in London's residential buildings at the middle super output areas (MSOA) and lower super output areas (LSOA) with multilayer neural network (MNN), multiple regression (MLR), random forest (RF), and gradient boosting (GB) algorithms, and factors related to socio-demographic, economic, and building characteristics were used as predictors. The results revealed that building characteristics, household income, and the number of households were the most important predictors of electricity and gas consumption. We also found that MNN models have outperformed MLR, RF and GB models in electricity and gas consumption prediction at MSOA and LSOA levels, with R2 values over 0.99 for the electricity consumption model. In summary, this study shows that the MNN models can be a useful tool to assist the formation of energy efficiency policies in buildings at MSOA and LSOA levels. • Four types of data-driven residential energy models at district scales in London were developed. • The performance of the developed models was validated against the independent testing data. • Multilayer neural network models outperformed the other three models. • Model inputs from UK census survey were effective in the model development. [ABSTRACT FROM AUTHOR]