In Search of Deep Learning Architectures for Load Forecasting: A Comparative Analysis and the Impact of the Covid-19 Pandemic on Model Performance

Abstract

In power grids, short-term load forecasting (STLF) is crucial as it contributes to the optimization of their reliability, emissions, and costs, while it enables the participation of energy companies in the energy market. STLF is a challenging task, due to the complex demand of active and reactive power from multiple types of electrical loads and their dependence on numerous exogenous variables. Amongst them, special circumstances-such as the COVID-19 pandemic-can often be the reason behind distribution shifts of load series. This work conducts a comparative study of Deep Learning (DL) architectures-namely Neural Basis Expansion Analysis Time Series Forecasting (N-BEATS), Long Short-Term Memory (LSTM), and Temporal Convolutional Networks (TCN)-with respect to forecasting accuracy and training sustainability, meanwhile examining their out-of-distribution generalization capabilities during the COVID-19 pandemic era. A Pattern Sequence Forecasting (PSF) model is used as baseline. The case study focuses on day-ahead forecasts for the Portuguese national 15-minute resolution net load time series. The results can be leveraged by energy companies and network operators (i) to reinforce their forecasting toolkit with state-of-the-art DL models; (ii) to become aware of the serious consequences of crisis events on model performance; (iii) as a high-level model evaluation, deployment, and sustainability guide within a smart grid context.

A comparative assessment of deep learning models for day-ahead load forecasting: Investigating key accuracy drivers

“…Subsequently, the MLR-based framework used for explaining the forecasting performance of the DL models is presented. The experimental process took place using an automated machine learning operations (MLOps) pipeline developed with MLflow [2], building up to the one described by Pelekis et al [44].…”

“…Wen et al [63] and Grabner et al [14] have also used variants of N-BEATS for probabilistic STLF and consumer-level STLF at global scale, respectively. In addition, Pelekis et al [44] compared the accuracy of TCN and N-BEATS in STLF settings for the case of the Portuguese national load at a 15-minute resolution, with N-BEATS resulting to superior forecasts. With respect to transformers, Zhang et al [66] evaluated a time augmented transformer for STLF on the electrical load data of New South Wales in Australia, Lim et al [33] validated temporal fusion transformers for STLF on the UCI Electricity Load Diagrams Dataset [60], while Huy et al [25] employed the latter architecture to forecast the load of Hanoi city using weather and calendar features.…”

“…As a result, this analysis provides valuable insights to energy stakeholders interested in reinforcing their forecasting toolkits with state-of-the-art DL techniques or producing forecasts in a selective fashion. (c) In addition to the above, we expand the analysis conducted by Pelekis et al [44] in three dimensions. First, we focus on purely autoregressive models, thus assessing the performance of DL time series forecasting models that do not require weather forecasts or any other type of external information as additional input.…”

“…This is an expected result as sNaive models simply replicate past values and do not involve any learning process. Last but not least, it is noticeable that although the validation set mainly contained a period that has been highly affected by the COVID-19 pandemic [44], the selected DL architectures manage to perform impressively well.…”

“…Additionally, the necessity for higher penetration of RES urges for more accurate forecasts that effectively adapt to the flexibility and demand response requirements posed within transmission and distribution operation systems. As a result, STLF has been investigated in a variety of research studies and projects [5,44,29,42,62].…”