Deep Learning Techniques for Load Forecasting in Large Commercial Buildings

Abstract

As large scale energy management strategies have gradually shifted the focus from the producer to the consumer side, buildings are starting to play a critical role in the efficient management of the electrical grid. Moreover some buildings have become prosumers by integrating local generation capabilities from renewable sources thus inducing additional complexity into the operation of the energy systems. As alternative to conventional energy consumption modelling techniques, a blackbox input-output approach has the ability to capture underlying consumption patterns and trends while making use of the large quantities of data being generated and recorded through dense instrumentation of the buildings. The paper discusses and illustrates an approach to apply deep learning techniques, namely Recurrent Neural Networks implemented by means of Long Short-Term Memory layers, for load forecasting. We focus on large commercial buildings which can be better managed by central operators and where better models can result in significant energy savings and broad economic and social impact. The case study is illustrated on two university buildings from temperate climates over one year of operation using a reference benchmarking dataset for replicable results. The obtained results show promise and can be further used in reliable load management algorithms with limited overhead for periodic adjustments and model retraining.