Generating scenarios from probabilistic short-term load forecasts via non-linear Bayesian regression

Abstract

In this paper we present a simple and intuitive method for fitting a non-linear Bayesian regression model on short-term load forecasts. Such models have been implemented via Bayesian neural networks, which are known for their hyperparameter sensitivity. We instead show a more general method to fit any regression model and demonstrate this by using a tree-model. Further, we evaluate the results against non-linear quantile regression, a common technique in probabilistic load forecasting. The resulting model allows to generate samples for future scenarios and thus can be applied to operations problems such as dynamic control of battery storage, an application that quantile regression is unfit for.

Low-complexity Risk-averse MPC for EMS

“…The second problem of interest, denoted P 2 , does not consider probabilistic chance constraints; hence, by setting η j|c → 0 in (12c), P 2 is reduced to a version similar to standard nominal deterministic MPC where we are only concerned in adhering to hard output constraints as formulated by (8). Probabilistic forecast models have been trained on historical power generation and load demand data, as obtained from the Skagerak Energilab pilot [15]. These Markov process models allow sampling a forecast prediction d k (ξ) for a given scenario ξ ∈ Ξ.…”

“…The former are of particular interest where one is typically concerned with multiple ESS in a microgrid setting. Although the computational complexity for solving the MSPE problem ( 15) is low, an interesting question would be how to reduce the number of historical samples used during this problem and instead use methods within reinforcement learning to adapt the underlying parameters between successive solutions of (15).…”

An improved LSTM-Seq2Seq-based forecasting method for electricity load

“…The results demonstrated that the seasonal regression method presented the best approximation effect, while a combination of the time series methods assisted in reducing the approximation error. Löschenbrand et al (2021) developed a fitted non-linear Bayesian regression model utilizing Bayesian neural networks for short-term load forecasting. It permitted the generation of samples from non-linear probabilistic forecasts that do not require the tuning properties of deep-learning methods.…”

EMS2aaS: A Dockerized framework for remote EMS deployment

“…In the context of this paper, the goal is to offer a cloud application (software delivered over the Internet) as a service to optimize the energy management of a particular microgrid. The EMS 2 aaS framework illustrated in this paper consists of three main parts: (i) A cloud-based EMS solution that scales for microgrid hardware-in-the-loop (HWIL), and computer-in-the-loop (CIL) simulations; (ii) Probabilistic load forecast modelling using non-linear Bayesian regression [3]; and (iii) Stochastic Model Predictive Control (SMPC) as EMS [4]. This paper addresses the work related to (i), providing a foundation for the implementation and the deployment of the parts at (ii) and (iii).…”

“…An objectrelational database includes the benefits of object-oriented approaches into a relational database [24]. This makes for a 2 For more information about the probabilistic load forecasting we refer to an accompanying paper also submitted for dissemination [3]. more powerful and advanced database with more possibilities for in-build objects and complex procedures, compared to a relational database like MySQL.…”