Constructing Flexible Time-series Forecast Models of Photovoltaics Using Advanced Machine Learning

One of the most exciting news from renewable energy sources is the breakthrough of photovoltaic energy, coupled with increasing demands for better energy management systems to integrate PV energy into the modern power grid, thus requiring efficient forecast methods for PV energy. Current open state-of-the-art machine learning models for long-term time-series forecasts for PV energy utilize classic transformer models. Here the paper develops a dynamic architecture of hybrid-transformer mechanism, empowering the model to learn from both power production and weather datasets more flexibly than previous architectures. Specifically, this paper discusses three hybrid-transformer models from the architecture-LSTM-based transformer, GRU-based transformer and VAE-based transformer, and interprets the traits of these models. The paper demonstrates these models' effectiveness by studying theoretically and validating experimentally through their train loss and test loss results. In the test results, the VAE-based transformer model proves the most compatible compared to LSTM-based and GRU-based transformers models. The paper's proposed architecture and specific interpretations can be used to address a wide range of scientific and technological applications.