Day-ahead composite load forecasting for more informed load modelling and network operation

• Information about the composition of aggregated demand improve network analyses. • Forecasting of load composition can help aggregators identify controllable loads. • Accuracy of artificial intelligence methods is area-specific. • At the aggregated level neural networks have similar performances. Demand side management (DSM) programmes offer promising solutions to facilitate operation and planning of modern power systems. A key aspect in successful deployment of DSM programmes is defining the level of flexibility of the demand side, i.e., the contribution of controllable loads to the total demand. Estimation or forecasting of load composition can help aggregators or suppliers identify controllable loads and times of the day when DSM programmes are more impactful. The information about load composition facilitates more comprehensive load modelling of the demand, which enables more accurate short-term and long-term network studies. This paper presents a methodology for day-ahead forecasting of the composite load model and flexibility of aggregated active and reactive demand in distribution network. The research relies on sub-metering data at end users’ premises and focuses on the application of feedforward artificial neural network (FFANN), long short-term memory (LSTM) and gated recurrent unit (GRU) for forecasting active and reactive loads, with subsequent decomposition of these loads into the composite load model components. As the accuracy of artificial intelligence (AI) methods tends to be area-specific, the proposed flexibility forecasting method was tested on the electrical demand of two cities in India, namely Bangalore and Itanagar. Load decomposition based on the forecasting results reveals that FFANN generally produces the most accurate predictions of load components, although LSTM and GRU provide comparable outcomes in specific cases. The study highlights the importance of selecting the appropriate neural network model and input configuration, with significant implications for improving forecasting accuracy and load management in power systems.