A Multi-layer Weather Classification-based Regression Model for PV Power Prediction

The intermittency of photovoltaic (PV) poses significant challenges to grid operation, where accurate PV power forecasting is crucial. Recent forecasting methods show low accuracy during days with rapid fluctuations of weather conditions. To overcome this, a novel multi-layer weather classification-based regression model (MWCR) is developed to predict the PV output power in this paper. The key innovation lies in classifying weather conditions at each time step based on irradiance, temperature, wind speed and humidity. However, due to the absence of irradiance sensors, the local irradiance on a clear sky cannot reflect the actual sky condition. To address this, our novel idea is to use the ratio of the predicted power obtained by LSTM over the maximum power in a clear sky calculated by the physical model to account the sky conditions. For each class, a regression model for PV power output is developed with local irradiance, temperature, wind speed, humidity and power ratio as inputs. The model is evaluated using realistic data collected from a 2-MW solar farm located at Rochester Institute of Technology. Results show that our model outperformed the single regression, conventional neural networks and gated recurrent unit and LSTM models for the testing and training data. In addition, MWCR achieves better results over LSTM on a summer and a winter day, with a significant improvement on the winter day.