Predictive Modelling of Electrical Loads for Grid Infrastructure Planning - Segmentation and Stochastic Simulation of Residential and Commercial Energy Demand with Machine Learning

Abstract

Abstract The transition towards a more sustainable energy system has increased the need for more precise and interpretable models for predicting future electricity demand. This thesis proposes a machine-learning based framework for predictive modelling of electrical loads, with the goal of enhancing grid infrastructure planning, particularly at the local level. The framework encompasses three components: (1) a load decomposition model that separates historical load into base load, temperaturedependent load and a solar generation component using only basic information such as temperature and solar radiation, (2) a building-informed segmentation model that estimates the coefficients for these load categories based on aggregated local building data, and (3) a residual process model that is used to model the stochastic variations that are not accounted for in the previous models. The models implement statistical methods, including machine learning with neural networks, modelling of stochastic processes and Monte Carlo simulations. The proposed framework can be used to provide several insights, such as peak demand, load duration curves and how these are affected by changing climate patterns or construction of new buildings. In addition to estimating extreme values, the framework can be used to model daily demand patterns, valuable when evaluating adoption of sustainable energy sources which are often less controllable. By explicitly linking electricity usage to observable variables such as temperature and solar radiation, the models allow energy and utility companies to make data-driven decisions for grid planning without sacrificing interpretability or operational transparency.