Multi-Time Instant Probabilistic PV Generation Forecasting Using Quantile Regression Forests

Abstract

Long-term planning for the reinforcement of power systems with PV-integration requires multi-time instant PV generation uncertainty modeling. Probabilistic forecasting of PV generation plays a vital role in the uncertainty management in power systems with PV penetration. An ensemble approach for probabilistic PV generation forecasting, such as the quantile regression forests, proves to be a suitable model because it models the uncertain PV generation more accurately compared to single mean models. The inherent nature of forests to prevent over-fitting by "bagging" the training data is an advantage. Also, the optimal choice of the model hyper-parameters adds to its efficiency as a forecaster. Further, the stochastic nature of weather conditions needs the selection of sensible regressors for the proposed quantile regression forests framework based on the physics of the underlying phenomenon. Real-world data for PV generation collected at multiple instants of time from the USA are employed to test the efficacy of the proposed probabilistic forecasting. The proposed model is compared against the basic quantile regression approach in terms of the accuracy of the quantile forecasts as well as prediction intervals using suitable scores and error metrics.