A Comparative Study on Prosumers Load Demand and Production Forecasting Using Machine Learning and Time Series Techniques

A data-driven and realistic policy could overcome the current turmoil in the intricate energy landscape. With solar and wind expanding faster than liquified natural gas or nuclear ever have, local, bottom-up solutions, such as organising collaboratively residential prosumers into electricity trading communities and the efficient use of aggregators, are one way to resilience and control over these resources. This paper analyses the features and characteristics of some prosumers and facilitates the identification of relevant algorithms for the specific tasks of both load demand and production forecasting using machine learning and time series techniques, thus enabling them to operate on a level playing field in the market. Load demand and production forecast simulations on two data sets, one collected in 2019, corresponding to 50 residential households and a public building from Portugal, and the other from 2023, that aggregates values from 29 household prosumers from Romania, show that both simple and double exponential smoothing techniques are more accurate than Markov predictors. Provided that the external contingencies influencing either the consumption or the production are considered, the latter gives a better prediction accuracy for energy production, whereas the former provides a better forecast for the consumption.