A self-adaptive deep learning model for building electricity load prediction with moving horizon

• Published in: Machine learning with applications Vol. 7; p. 100257
• Main Authors: Luo, Xiaojun, Oyedele, Lukumon O.
• Format: Journal Article
• Language: English
• Published: Elsevier Ltd 15.03.2022; Elsevier
• Subjects: Deep learning; Deep neural network; Moving horizon; Particle swarm optimisation; Prediction; Self-adaptability; Deep learning; Moving horizon; Deep neural network; Self-adaptability; Particle swarm optimisation; Prediction
• ISSN: 2666-8270; 2666-8270
• DOI: 10.1016/j.mlwa.2022.100257

A self-adaptive deep learning model powered by ranking selection-based particle swarm optimisation (RSPSO) is developed to predict electricity load in buildings with moving horizons. The main features of the load prediction model include its self-adaptability, repeatability, robustness and accuracy. In real-world building applications, the relationship among weather data, time signature and electricity load is quite complicated. In the proposed self-adaptive deep learning model, a deep learning model with multiple hidden layers is implemented to improve prediction precision. Meanwhile, RSPSO is implemented to select the network’s optimum architecture, which involves discrete variables (i.e. the quantity of neurons in each layer and the quantity of hidden layers) and categorical variables (i.e. activation function in each layer and learning approach). Moreover, the moving horizon approach is adopted to update the architecture and structure of the dynamic deep learning model while enabling its capability in capturing the latest featuring patterns in the electricity load of the building. The proposed load prediction model is tested with the local meteorological profile and electricity load of an educational building. The self-adaptive load prediction model is identified to be the most effective at forecasting the next horizon’s energy consumption, while its prediction performance would deteriorate with the increase of time. The mean squared error, mean absolute error, and coefficient of determination of the proposed prediction model are within the range of 4.48 kW–11.23 kW, 1.28 kW–2.31 kW and 97.52%–98.92%, respectively, demonstrating its prediction accuracy and repeatability. When Gaussian white noise is added to meteorological data, the increase in mean absolute error is within the range of 2.08%–15.33%, demonstrating the robustness of the proposed prediction model in overcoming uncertainty in the weather forecast. Therefore, the proposed accurate, robust, repeatable and self-adaptive load prediction model can be rooted in practical energy management systems thus facilitate building operation and system control. [Display omitted] •A ranking selection-based particle swarm optimisation (RSPSO)-enhanced self-adaptive deep learning prediction model.•RSPSO for determining discrete and categorical design variables for optimal architecture of deep neural network.•Moving horizon make architecture and structure self-adaptive to cover the latest patterns in building energy consumption.•Actual weather data and measured energy consumption data are adopted as input datasets.•The proposed prediction model is accurate, repeatable, robust and self-adaptive.