Short-term apartment-level load forecasting using a modified neural network with selected auto-regressive features. (1st April 2021)

• Record Type: Journal Article
• Authors: Li, Lechen; Meinrenken, Christoph J.; Modi, Vijay; Culligan, Patricia J.
• Is Part Of: Applied energy. Volume 287(2021)
• Publication Date: 2021-04-01
• Subjects: Short term load forecasting -- Dynamic feature selection -- ConvLSTM neural network -- "Default" state of model -- Time-series electricity forecasting difficulty analysis; Power (Mechanics) -- Periodicals; Energy conservation -- Periodicals; Energy conversion -- Periodicals; 621.042
• DOI: 10.1016/j.apenergy.2021.116509
• Languages: English
• ISSNs: 0306-2619

Abstract

Highlights: Forecasted hourly apartment electricity use (59 apartments, 11 floors, 3 seasons). Developed novel neural network with selected auto-regressive features. Compared accuracy of model against 4 benchmark models. Novel neural network improves forecasting accuracy by up to 25%. Identified underlying drivers of accuracy and discussed their application. Abstract: Residential electricity load profiles and their diversity have become increasingly important to realize the benefits of Smart or Transactive Energy Networks (TENs). An important element of TENs will be practical, accurate, and implementable residential load forecasting techniques. While there have been many approaches to short-term load forecasting, few have included forecasting for individual households, partly because the high volatility and idiosyncrasies present in individual household load data can pose significant challenges. In this study, we develop a Convolutional Long Short-Term Memory-based neural network with Selected Autoregressive Features (termed a CLSAF model) to improve short-term household electricity load forecasting accuracy by employing three strategies: autoregressive features selection, exogenous features selection, and a "default" state to avoid overfitting at times of high load volatility. We include aggregations of apartments to floor and building level, because utilities may favor transactive approaches that rely on aggregator models, e.g., a cluster of consumers as opposed to an individual. We demonstrate that the CLSAF model, by virtue of its enhanced feature representation and modest computational resources, can accomplish load forecasting in a multi-family residential building across three spatial granularities (individual apartment/household, floor, and building levels), with an accuracy improvement of up to 25% compared to a persistence model. We propose a data screening technique to characterize time-series electricity-load data. This technique is suitable for integration into a TEN ecosystem and allows one to estimate confidence levels of the load forecasts to optimize computational resources and the risks associated with uncertain forecasts.