Probabilistic PV Power Forecasting by a Multi-Modal Method using GPT-Agent to Interpret Weather Conditions

Traditionally Photovoltaic (PV) power generation forecasting is based on numeric meteorological vectors to capture weather conditions, which generally misses valuable multimodal information such as those contained in linguistic weather descriptions. In the light of the recent success of large language models (LLM), this paper presents a new method for probabilistic PV power forecasting that integrates a generative pretrained transformer (GPT) agent to interpret the linguistic descriptions of weather. The GPT-Agent interprets the descriptions of weather conditions from a specialized API such as Openweather and transforms them into estimated numeric vectors representing the likelihood of cloud coverage. Then, a compact multimodal feature input is constructed combining numeric meteorological data and the interpreted weather conditions. Using this multimodal input, two XGBoost models are hierarchically trained for probabilistic PV irradiance and power forecasting. With the first layer of XGBoost model to forecast the expected value of the PV irradiance and second layer of XGBoost model to forecast its variance, a probabilistic irradiance interval is constructed, and subsequently the PV power generation can be calculated and probabilistically forecasted. The multimodal approach offers a holistic perspective on environmental factors affecting PV generation, resulting in more accurate results. Furthermore, the GPT -agent offers interpretability of PV forecasting by indicating the reasons of forecasting the PV output based on the weather. Simulation testing results demonstrate the effectiveness and advantage of the proposed method.