Machine learning-based short-term net load forecasting for residential buildings with integrated photovoltaic systems

Short-term net load forecasting (STNLF) is crucial for the efficient operation and management of residential buildings integrated with photovoltaic (PV) systems. However, the inherent variability of PV generation introduces significant uncertainty, leading to increased forecasting errors. To address this challenge, this study proposes a machine learning model based on Bayesian neural networks (BNNs) for direct STNLF in solar-integrated residential buildings. BNNs were selected due to their high forecasting accuracy under highly variable operating conditions. Historical data from households with rooftop PV systems in Cyprus were used for the model evaluation procedure. As such, the BNN model was tested across individual and aggregated households at 30- and 60-minute resolutions and benchmarked against a reference naïve persistence model (NPM). The BNN model achieved normalized root mean square error (nRMSE) values below 10% across all investigated households and resolutions. Furthermore, it recorded skill scores of up to 49.18% over the NPM. Moreover, the BNN model consistently outperformed the NPM under various irradiance conditions, demonstrating improved accuracy and robustness. Overall, the results prove the superiority of the BNN model and its potential for integration into grid operators' decision-making tools/platforms to enhance residential energy management.