ELECTRIC LOAD AND POWER FORECASTING USING ENSEMBLE GAUSSIAN PROCESS REGRESSION

Accurate week-long forecasting of load demand and generation scheduling is critical for efficient operation power grid systems. In this work we present an ensemble Gaussian process regression (EGPR) method week-ahead periodic time series data. The proposed EGPR based on the GPR method, employs constructed by windowing data to compute prior mean covariance. To improve estimates statistics from a potentially small avoid rank-deficiency issues, propose leave-one-out cross-validation shrinkage approach regularizing covariance estimates. Furthermore, evaluate existing available in literature. A synthetic set describing dynamics with 700 buses 134 generators real total system Duke Energy Ohio are used test method. Both sets contain collected every hour over 365-days period. also contains profiles each generator. We demonstrate that capable accurately weekly outperforms traditional methods, including standard data-driven GPR, autoregressive integrated moving average (ARIMA), TBATS (Exponential smoothing state space model Box-Cox transformation, ARMA errors, Trend Seasonal components) methods.