Deep learning and transfer learning models of energy consumption forecasting for a building with poor information data

Abstract

Abstract Precise prediction of energy consumption in buildings could significantly optimize strategies for operating building equipment and release the energy savings potential of buildings. With advances in computer science and smart meters, data-driven energy forecasting models, particularly deep learning models, are becoming increasingly popular and can achieve good prediction accuracy. However, these models require a multitude of historical data from predicted buildings for training, which are difficult to acquire for newly constructed buildings or buildings with newly established measurement equipment. In order to obtain satisfactory prediction accuracy under such poor information state, this paper proposes two deep learning models, which are a sequence-to-sequence (seq2seq) model and a two dimensional (2D) convolutional neural network (CNN) with an attention layer, and transfer learning framework to improve prediction accuracy for a target building. A case study of three office buildings is discussed to demonstrate the proposed method and models. Compared with the results of a long short-term memory (LSTM) network with poor information state, the seq2seq model improved forecast accuracy for a building with a small quantity of data by 19.69 percentage points in mean absolute percentage error (MAPE), and the 2D CNN model by 20.54 percentage points, on average.