A method for generating radiation meteorological year using machine learning fusion and adaptive neural network with application to BIPV systems

The Radiation Meteorological Year (RMY) is a specialized dataset developed to accurately characterize regional solar radiation conditions, thereby facilitating the precise design and performance evaluation of photovoltaic (PV) systems. With the rapid advancement of PV technologies and their increasing deployment, there is a corresponding demand for high-resolution meteorological data that can reflect both spatial and temporal variations. However, traditional Typical Meteorological Year (TMY) generation methods often rely on fixed-weight statistical schemes, which inadequately capture the temporal dynamics and periodic patterns inherent in solar radiation data. This limitation is particularly critical in regions exhibiting complex climatic behavior, where traditional TMY may introduce significant biases in PV performance modeling. To address this gap, the present study proposes a novel RMY construction method based on a Bidirectional Long Short-Term Memory (Bi-LSTM) neural network. By learning temporal dependencies in both forward and backward directions, the model autonomously selects month sequences that align most closely with long-term solar radiation trends. A case study involving a 12-story building-integrated photovoltaic demonstrates that the proposed approach significantly outperforms benchmark methods—namely the Sandia method, the Typical Principal Component Year (TPCY), and a recent machine learning-multimodal fusion (ML-MF) approach—achieving up to a 5.6 % improvement in energy consumption prediction accuracy and a 5.0 % reduction in PV generation error. These findings underscore the method's potential for generating high-fidelity, regionally adaptive meteorological datasets for advanced renewable energy applications.