Scalable and Lightweight Machine Learning Based Load Forecast: Netload versus Disaggregrated Forecast
Paper i proceeding, 2024
This paper develops lightweight and adaptive demand forecast models for a residential building integrated with solar photovoltaics using scalable and adaptive deep learning algorithms, i.e., long short-term memory (LSTM) and gated recurrent units (GRU). First, the forecast models have been trained using the real measurement data from a residential building. Then, the models have been used in case studies using the real-time data for two forecasting approaches: i) netload forecast; ii) disaggregated forecasts, i.e., forecasting the load and PV generation separately. The performance of the two forecasting approaches have been compared. The results from case studies showed that disaggregated forecast approach was superior (with an overall RMSE of 2.03 kW for the building with max demand of 10.53 kW) than the aggregated forecast approach (with an overall RMSE of 2.63 kW). Case studies results have also demonstrated that the models are scalable with more data, and are lightweights, hence, suitable for resource-constraint devices. Although LSTM shows advantages in accuracy, GRU shows better scalability in terms of computational efficiency. The models can be utilized by various stakeholders, such as building owners, grid operators, etc., and can be adapted to other types of buildings.
forecasting
lightweight
scalability
deep learning
netload
energy optimization
neural networks