Residential baseload-forecasting by applying recurrent neural networks with gated recurrent units on field data

Buildings account for a significant share of the overall energy consumption. Currently, their components are becoming increasingly intelligent and controllable. However, a significant proportion of energy consumption remains uncontrollable. Accurate forecasting of this uncontrollable portion is crucial for effective system planning. To address this, forecasting models based on the gated recurrent unit (GRU) architecture were trained using 52 real-world datasets. The GRU architecture is known for delivering high accuracy while maintaining lower complexity than similar architectures like Long Short-Term Memory (LSTM). Model evaluations indicate superior accuracy compared to persistence models, with further accuracy gains observed as data resolution and aggregation levels increase.