To integrate increasing amount of renewable energy sources, an accurate prediction of power generation on a fine temporal and spatial scale is critical. It will fundamentally enable various use cases such as efficient scheduling and stable operation of a transmission grid, an economical bidding in the energy market, generation-aware maintenance planning, and optimal operation of the grids. However, the requirement of more granular generation forecasting for individual wind turbines and small utility-scale solar farms (i.e., O(1MW) imposed significant constraints on numerical weather prediction.
To address this local solar and wind generation forecasting capability, we use a state-of-the-art high spatio-temporal resolution weather forecasting engine, Deep Thunder. The forecasting engine is customizable to meet the needs of specific weather-sensitive business decisions. It is based, in part, on the ARW core of the Weather Research and Forecasting (WRF) model. The weather model generates variables including both direct model output as well as diagnostic fields derived from specialized post-processing. These data then permit execution in parallel of data-driven (i.e., via statistical and machine-learning) models to predict wind and solar power. All of these models operate at a granularity that enables aggregation from the 1 km computational grid of the weather model with up to three-day lead time.
The core of the renewable power forecasting method exploits machine learning techniques that adaptively correct temporal bias and regression bias from weather models. Specifically, we use historical power production data to estimate temporal bias and regression bias. The method uses a variation of a finite impulse response (FIR) filter. The FIR filter is adaptively tuned using a numerical optimization which minimizes a desired error metric. We use the mean absolute error of the power estimation as an objective term.
The proposed method is adaptively tuned using the running 28 to 45 day historical power measurement and weather forecasting data. This feature captures the variation of the temporal bias and regression bias that change season to season. In addition, a separate set of parameters are being tuned for each forecasting cycle (0Z and 12Z) as well as for each day cycle (0-24, 25-48, and 49-72 hours).
Our results present a forecasting accuracy of 7.3% normalized mean absolute error for solar farms and 10% normalized mean absolute error for wind farms on average. We will present the approaches, lessons learned, and outcomes from four wind farms (119 MW in total capacity) and 21 utility-scale solar farms (39MW in total capacity) in Vermont.