Learning With Correlation-Guided Attention for Multienergy Consumption Forecasting

With the advent of advanced metering infrastructure (AMI) technologies, various energy sources, such as gas, heating, and water can be actively collected. In this study, using multienergy sources, we aim to improve a prediction model for the consumption of a target energy source by exploiting the inherent correlations with other energy sources, which is a unique feature of the problem in this study. To achieve this, we propose a learning model based on a correlation-guided attention mechanism. We design our model with a two-stage learning strategy, wherein the model learns through two kinds of distinct loss functions, effectively capturing the correlations among energy sources and incorporating the learned weights into the prediction model. Through extensive experiments using real-world datasets, we demonstrate the effectiveness of our model based on six distinct types of neural network-based models while varying target energy sources, the used energy sources, and datasets.