Adaptive sliding window load forecasting

Abstract

Small-scale, renewable generation which is embedded in the distribution network is causing previously unseen fluctuations in demand. In Northern Ireland this new generation, which is not visible to, or controllable by, the system operator, is presenting major challenges for accurate load forecasting. Currently deployed load forecasting methods are struggling to cope due to the rapid growth in this new generation, and its weather dependent nature. In this paper linear load forecasting methods are investigated within a sliding window parameter updating framework, which is adopted to address the nonstationarity of the problem. Initially, models are built using historical load terms selected based on correlation analysis of recorded load data. Then, Forward Selection Regression is used to choose the most important variables from a candidate set, consisting of historical load variables and a range of weather related parameters. Model performance is evaluated on load data for the period 2015-2016. A 7-input model, with parameters updated on the basis of a 5-day sliding window of historical data, is shown to yield optimal results, with a mean absolution percentage error of 2.4%.

Simultaneous Estimation of Behind-the-Meter Solar and Wind Power at the Bulk Supply Point

“…solar capacity, weather and calendar variables) to net demand variations. Foster et al investigated a similar strategy using a forward selection regression technique [9] for linear models and correlationbased approach [10] for nonlinear models to select the most appropriate variables. Wang et al proposed a method in [11] to model the conditional forecast residual and use it to improve the results of probabilistic forecasting methods.…”