Low-Complexity Neural Networks for Electrical Load Forecasting in Renewable Energy Communities

The problem of electrical load forecasting represents a crucial aspect in many Smart Grid applications. In Renewable Energy Communities, an effective Energy Management System, aiming at improving clean energy consumption and energy self-sufficiency, schedules the operations of the available controllable loads and energy storage systems necessarily relying on load forecasts. Obtaining accurate load predictions is a particularly challenging task, given the randomness inherent in the electrical load time series. As such, advanced Artificial Intelligence and Deep Learning techniques may result ineffective if additional information (e.g., external temperature) is not provided. Since these data are often not accessible, this paper studies the performance of low-complexity models, based on Feed-Forward Neural Networks, applied to the forecasting of the mere electrical load signal. This kind of approach in the context of an Energy Management System presents two major advantages: first, fast-trainable neural models can be retrained online to adapt to new consumption habits, without affecting the operations of the Energy Management System; second, a short inference time allows for having more time for other demanding operations, namely the communication with metering and actuating devices and the elaboration of the control strategy. The results on an open-source electrical load database show that the proposed models can achieve a forecasting accuracy comparable to recurrent networks, but with shorter training and inference time.