Applied Machine Learning for Short-Term Electric Load Forecasting in Cities - A Case Study of Eskilstuna, Sweden

Abstract

With the growing demand, electrification, and renewable proliferation, the necessity of being able to forecast future demand in combination with flexible energy usage is tangible. Distribution network operators often have a power capacity limit agreed with the regional grid, and economic penalties await if crossed. This paper investigates how cities could deal with these issues using data-driven approaches. Hierarchical electric load data is analyzed and modeled using Multiple Linear Regression. Key calendar variables holidays, industry vacation, ”Hour of day” and ”Day of week” are identified alongside the meteorological heating-, and cooling degree hours, global irradiance, and wind speed. This inexpensive algorithm outperforms the benchmark ”weekly Naïve” with a relative Root Mean Squared Error of 35% for the year-long rolling origin evaluation. Learnings from the data exploration and modeling are then used to evaluate the AI-based model Light Gradient Boosting Machine. Using similar explanatory variables for this expensive algorithm results in a relative error of 45%, although it outperforms the previous one during the summer. The models have varying strengths and weaknesses and could advantageously be combined into an ensemble model for improving accuracy. Incorporating detailed knowledge of local renewable electricity production in combination with hierarchical forecasting could further increase accuracy. With domain knowledge and statistical analysis, it is possible to create robust load forecasts with acceptable accuracy using easily available machine-learning libraries. Both models have good potential to be used as input to economic optimization and load shifting.

Deriving Input Variables through Applied Machine Learning for Short-Term Electric Load Forecasting in Eskilstuna, Sweden

“…The change in accuracy of the models when each of the explanatory variables was consecutively added is documented in Table 1. The results for the 2021-2022 year were gathered from previously published work [39].…”