A comparative analysis of artificial neural network architectures for building energy consumption forecasting

Authors	Jihoon Moon, Sungwoo Park, Seungmin Rho, Eenjun Hwang
Title	A comparative analysis of artificial neural network architectures for building energy consumption forecasting
Publication	International Journal of Distributed Sensor Networks
Volume	15
Issue	9
Pages	x
Year	2019
DOI	https://doi.org/10.1177/1550147719877616

Introduction

Background

• Smart grids are a solution for energy shortage and environmental pollution, providing reliability, flexibility, sustainability, and efficiency.
• A smart grid is a platform for exchanging real-time power information between suppliers and consumers.
• Typical smart grids include smart meter, EMS (Energy Management System), ESS (Energy Storage System), and diverse RES (Renewable Energy Source).

Previous Research

• EMSs collect and analyze data related to building energy consumption to save energy on the demand side and generate schedules for power generation and ESSs on the supply side.
• Accurate STLF (Short-Term Load Forecasting) is required to generate more effective schedules.
• STLF is used to predict the electric load on an hourly basis up to 1 week in advance.
• Accurate STLF can provide economic benefits by storing energy at night when electric costs are relatively low and emitting electricity during the day when electric costs are high.
• STLF is a challenging task due to complex energy consumption patterns and uncertain external factors.

Proposed Model

• This study constructs an ANN (Artificial Neural Network)-based STLF model for accurately forecasting electric energy consumption of a building or building clusters.
• General factors such as calendar data, weather information, and historical electric loads are considered for applying the STLF model as the baseline model for other buildings or building clusters.
• The model predicts the 30-min interval electric load for five different types of buildings by setting several cases as test sets.
• The performance of various activation functions and number of hidden layers is extensively compared to construct an optimal ANN-based STLF model.

Significance

• The study aims to improve the performance of STLF in a smart grid context, which is critical for ensuring the reliability of the electric power system equipment and preparing for losses caused by power failures and overloading.
• The study's primary contributions are to build an accurate ANN-based STLF model, consider general factors for applying the model to other buildings or building clusters, and extensively compare the performance of various activation functions and number of hidden layers for constructing an optimal ANN-based STLF model.

Proposed Model

• An ANN or MLP (Multi-Layer Perceptron) is a ML (Machine Learning) algorithm that has an FFNN (Feed Forward Neural Network) architecture with an input layer, one or more hidden layers, and an output layer.
• The number of layers and nodes, as well as the activation function, affect the performance of the network, and the number of hidden layers determines the depth or shallowness of the network.

• ReLU (Rectified Linear Unit) has been consistently used as an activation function when the number of hidden layers is two or more, but using ReLU can result in deactivated neurons and slow learning.
• Other activation functions, including LReLU (Leaky Rectified Linear Unit), PReLU (Parametric Rectified Linear Unit), ELU (Exponential Linear Unit), and SELU (Scaled Exponential Linear Unit), have been introduced to solve these problems and have been used for constructing STLF models.
• ANN-based load forecasting models are constructed using all possible combinations of hidden layers from 1 to 10 together with activation functions, including ReLU, LReLU, PReLU, ELU, and SELU.
• The number of hidden neurons is determined to be 2/3 the size of the input layer plus the size of the output layer, and 81 nodes are used for the forecasting model.
• Xavier initialization is used to sort initial weights for individual inputs in a neuron model, and the learning rate and learning epoch are important hyperparameters to be considered.

Experiment

Average CVRMSE

Average MAPE

• SELU exhibits the best performance, and ELU exhibits the worst performance in most cases.
• SELU's superior self-normalization quality enables it to be trained faster and better than other activation functions.
• ANN models with SELU repeatedly exhibit a higher frequency than other activation functions.
• ANN models with one hidden layer generally have poor predictive performance.

Average Ranking value of number of hidden layers for SELU

• SELU models with five and six hidden layers exhibit the lowest average ranking values of CVRMSE (Coefficient of Variation of the Root Mean Square Error) and MAPE (Mean Absolute Percentage Error), indicating excellent prediction performance.
• ANN models with one hidden layer generally have poor predictive performance.

CVRMSE,MAPE values when the statistical techniques and ANN with SELU and five hidden layers are used

• Compared to other statistical techniques such as Persistence, MA (Moving Average), ES (Exponential Smoothing), and MLR (Multiple Linear Regression), the ANN model with SELU and five hidden layers provides the best predictive performance.