Environmental & Climate Sciences Department Seminar

This seminar introduces our efforts to develop an integrative renewable energy forecasting system by combining physics-based Weather Research and Forecast (WRF) model and data-driven machine learning models, with a focus on wind forecast. Long-term surface wind simulations with WRF are first evaluated against observations at 22 sites over the New York State divided into six terrain types (i.e., continent, lakeside, river-valley, Long-Island, offshore-island and offshore-ocean). The results show that WRF model overestimates typical inland (i.e., continent) winds by 1 m/s throughout the whole day. WRF model underestimates diurnal variability of local circulation impacting surface wind and misses the physical nature that local circulation increases typical inland surface wind speeds. Simulated offshore-island winds cannot reproduce the observed diurnal variation but exhibit constant average and standard deviation; but offshore-ocean winds are successfully reproduced. Among the six terrain types, the WRF produces the best estimation of offshore-ocean winds, the secondly best estimation of Long-Island winds, and almost same performance in wind estimation at the other sites. The simulation of coastal region wind profiles shows remarkably overestimated Planetary Boundary Layer (PBL) wind speeds and underestimated free troposphere winds, whereas the WRF simulation at a mountain region site shows underestimated PBL wind speeds. Model deficiencies and potential reasons underlying the wind biases and the discrepancies among different terrains are explored. Comparative analysis reveals that the best data-driven machine learning model outperforms the physics-based WRF forecast in short lead time; but the WRF model achieves higher accuracy once the lead time is beyond a few hours.