LAUSR

Accurate short-term load forecasting (STLF) is required for reliable power system operations. Nevertheless, load forecasting remains a challenge owing to the high dimensionality and volatility of electrical load data as time series. In this study, a feature extraction framework for electrical load and other complementary data as a multivariate time series is proposed. The proposed framework consists of tagging and embedding processes that extract patterns from the multivariate time series as tags and capture their temporal and dimensional relations. In the embedding process, a network model that embeds the tags is deliberately designed with a convolutional layer in a multi-output structure based on mathematical analysis. Furthermore, a deep learning-based STLF model is constructed with the proposed feature extraction framework. The performance of the proposed STLF model for day-ahead load forecasting is evaluated on a publicly available set of real electricity demand data. The experimental results verify that the proposed approach reduces the root mean squared error by 5% to 12%. This improvement in load forecasts can benefit power grid operations as it provides more accurate expectations on the behaviors of the power grid in short term, which can be utilized in power grid applications, such as power dispatch and scheduling.