Bayesian probabilistic photovoltaic power forecasting and stochastic model-predictive control for an agricultural microgrid

This paper describes a novel probabilistic forecasting method for photovoltaic power for use in energy man agement for an agricultural microgrid. The forecasting method utilizes recent historical data and a general weather forecast to fit a spline + Gaussian process (GP) model using no-u-turn sampling (NUTS) to infer model parameters in a Bayesian modeling framework. The method seamlessly transitions from a near-term forecast dominated by recent output to a regime where output is dominated by the meteorological forecast. The forecasts are evaluated using proper scoring rules for multivariate probabilistic forecasts and are compared to a refer ence multivariate persistence forecast and to an LSTM-based forecasting method. The probabilistic method is integrated into a simulation of stochastic model-predictive control (SMPC) for an off-grid agricultural microgrid incorporating photovoltaic generation, a battery storage system, irrigation pumping, and local electrical loads. A 20 %-35 % reduction in simulated operating cost is achieved using the probabilistic forecast compared to a simple expected-value point forecast.