Lightweight Deep Learning for Photovoltaic Energy Prediction: Optimizing Decarbonization in Winter Houses

This paper proposes an innovative hybrid multivariate deep learning approach to predict photovoltaic (PV) energy production in
winter houses, with a focus on lightweight models with low environmental impact. A methodology is developed to assess the
carbon footprint of these models, considering training energy consumption, operational CO2 emissions, and energy savings from
PV production optimization. This approach allows selecting models that offer the best trade-off between predictive accuracy and
environmental responsibility. The study compares the performance of long short-term memory (LSTM), convolutional neural
networks (CNN), and a hybrid CNN-LSTM model for short-term PV production prediction in high-snow regions, using a Positive
Energy Winter House (PEWH) case study in Poschiavo, Switzerland. The results show that PV integration can reduce primary
energy consumption by up to 63%, with a decarbonization rate of 11%. However, full fac¸ade coverage leads to overproduction due
to limited winter sunshine and relatively low energy consumption. LSTM optimization identifies configurations (south facade or
north roof) achieving decarbonization rates of 131% and 116% respectively, covering 95% to 114% of energy needs, and limiting
overproduction. The PEWH case study demonstrates the potential of lightweight deep learning for optimized energy prediction and
decarbonization of buildings, especially in cold regions, and highlights the importance of the carbon impact of models in the face
of the increasing availability of PV data for more efficient and eco-responsible predictions.