Juan Luis Bosch, Chi Chow, Mohammad Jamaly, Matt Lave, Ben Kurtz, Patrick Mathiesen, Andu Nguyen, Anders Nottrott, Bryan Urquhart, Israel Lopez Coto, Handa Yang, Oytun Babacan, Ryan Hanna, Iman Gohari, Guang Wang, Ozge Serifoglu Mechanical & Environmental Engineering, http://solar.ucsd.edu Jan Kleissl, Associate Professor, Center for Renewable Resources and Integration, University of California, San Diego WRF SOLAR FORECASTING
Project Sponsors and Partners CPUC CSI Solar RD&D Program www.calsolarresearch.ca.gov
Motivation and Tools Goal: Reduce reserve requirements and integration costs. Horizon: > Intra-hour (5 min dispatch, 15 min pre-dispatch) > Day ahead: unit commitment Tools: Sky imagery, Stochastic Learning, NWP
Time Space and Time Scales of Solar Forecasts Tool: 1 day Sky Clouds there +- 100 ft, +-3% Imager Satellite Clouds exist +- 2 miles, +- 10% NWP Clouds likely CAISO 1 hour 10 min EPRI 1 min FPL 10 sec 1 sec Distance Point Yards / meters km / miles 10 miles 1000 miles
June 2010, 0630 PST Clear Sky Index SolarAnywhere WRF Marine Layer Solar Forecasting Cloud Evaporation Time -Evaluation period May October 2013
Marine Stratocumulus - Mechanisms More stratocu with: Stronger inversion Cooler SST Enhanced cold air advection Bennetts, D., McCallum, E., and Nicholls, S., 1986. Stratocumulus: An introductory account. The Meteorological Magazine. 115, 65-76 Myers, T. A., and J. R. Norris, 2013: Observational evidence that enhanced subsidence reduces subtropical marine boundary layer cloudiness. 6 J. Climate, in press.
High-resolution, cloud-assimilating NWP at UCSD: day-ahead (40 h) solar forecasts across Southern California. 2.5 km and 15 min resolution.
High-Resolution WRF is used to improve forecast accuracy 1) High-resolution to resolve small scale clouds and coastal gradients Domain Resolution (km) Points Size 1 12.5 125 x 125 1500 x 1500 km 2 2.5 4 124 x 124 496 x 496 km 3 1.3 100 x 100 133 x 133 km 2) Vertically dense spacing near surface (50 pts below 3 km) 3) Physics options chosen to promote marine layer clouds Physics Option Microphysics Radiation - LW Radiation - SW PBL Cumulus Land-Surface Thompson RRTMG Dudhia MYNN Kain-Fritsch Noah 8
Model and Cloud Data Assimilation Observations Cloud Analysis System WRF Model Initialization - Pre-processing - Ingest NAM To improve accuracy, clouds are populated in initial conditions GOES Imagery Direct Comparison Cloud Contingency Table Cloud Building/Clearing Spin-Up Time (2 hrs) Initial Cloud Fields Direct Cloud Assimilation - Cloud hydrometeors are modified in initial conditions WRF model simulates cloud evolution - Evaporation - Deformation - Condensation Updated Model Fields - q Vapor RUC methodology - Benjamin, et al., 2004, 2007, etc. 40 h forecast
GOES Δq v NAM B.C. Satellite Imagery Direct Cloud Assim. WRF Model Temperature Inversion Irradiance Forecast
WRF-CLDDA Simulation of Marine Layer Dissipation 6/13/2011
6/13/2011 Simulated Marine Layer Evolution by Cross Section
Forecast Performance SolarAnywhere (SAW) Analysis - Provides spatial coverage (0.01 ) of irradiance - Validated against NAM and WRF - Nearest neighbor interpolation to WRF grid Slight under-prediction Offshore (too thick of clouds) rmae is highest along coast - WRF is less biased/more accurate - Small coastal strip (~10-20 km has highest error)
TOD Error Dependence: Error is largest for early morning hours NAM WRF-CLDDA rmbe Large improvement over the ocean Coastal Strip WRF Error NAM Error 14
Joint PDF of Clear Sky Index at Coastline Observed NAM WRF kt* Joint-PDF * kt GHI GHI CSK WRF-CLDDA matches observation Error Intra-Day Day-Ahead (% GHI CSK ) Pers NAM WRF- NA WRF- CLDDA Pers. NAM WRF- NA WRF- CLDDA - Predicts ML reformation rmbe 3.3 17.8 6.5 0.4 2.2 14.4 2.6 1.7 Error rmae 22.4 25.4 21.3 21.3 28.8 24.0 18.4 18.2 rrmse 30.4 33.8 31.6 31.0 37.1 33.1 27.8 27.6 rstderr 30.2 28.7 30.9 31.0 37.0 29.8 27.7 27.5
Error Regime A: Inaccurately characterized ML ML Dissipation Irradiance over-predictions
sunrise ML Dissipation Occurs - T increases after sunrise - Vertical mixing/dissipation - Either T is too high or q v is too low Inversion is simulated as too strong and at a low altitude - Less water added in assimilation Height (m) Strength (K) Accurate ML Days Inaccurate ML Days -293 0.9-673 2.3
Error Regime B: Reverse Clearing-Type Days Irradiance underprediction
- Characterized by high night-time temperature inversions - Easy to predict occurrence - DA method invalid WRF - Cool temperatures aloft - Strong onshore (westerly) flow WRF-CLDDA MBE (W m -2 ) 54.6 12.7 MAE (W m -2 ) 124.1 128.2 - Orographically lifted clouds are too thick in model
WRF ENSEMBLES FOR SOLAR FORECASTING
Generating Variability in ML Cloud Burn-off Times Most NWP are biased dry and fail to predict daytime stratocu over land. Evaluate sensitivity of GHI forecasts: Multi-model ensembles: various combinations of microphysics, pbl, radiation, cumulus schemes. Multi-parameter ensembles: vary mixing length, heat transfer coefficients etc. > Mixing length within the PBL determine exchange of heat and moisture > alp parameters determining the dominant mixing length alp1: surface layer alp2: pbl alp3: top of pbl/entrainment Finding: Moisture Flux > Entrainment Rate > Turbulent Mixing 21
Sensitivity of coastal cloud to mixing length - alp1, alp2, and alp3 were changed to extreme values - Change in total water content, temperature, and clouds was observed - Single day 6/15/2011 typical marine layer type day - No cloud assimilation - All heat plots are deviations from RAP General Conclusions - As alp1 or alp2+alp3 increased, the cloud cover persisted longer over the coast - Together the effect was largest Increasing alp1/alp2/alp3 delays the dissipation of cloud cover and reduces irradiance 22
Increased alp2/alp3 Increased alp1 time Q cloud deviation from RAP - In general clouds are thicker inland for increased alp - Extend further inland - Higher water content Slight delay (more gentle slope to coast) z = 175 m z = 700 m Time Evolution of Cross section - Clouds slightly delayed in burning off - Likely due to higher water content 23
Total water content (q Cloud + q Vapor + q Rain ) deviation from RAP Increased q t over water Increased alp2/alp3 Reduced q t over land Increased alp1 Within PBL - Alp1 has a larger impact on q t than alp2/alp3 - Greater moisture transport from ocean [1] z = 175 m z = 700 m Large difference between land and water - Increased transport from water surface helps sustain coastal clouds - Water is generally reduced over land - At times, advection moves cloud water over land, helping delay the dissipation (at 1 in plot) 24
Conclusions of this sensitivity analysis - Increasing alp1 or alp2+alp3 will increase cloud cover in coastal areas - Clouds with higher water content - Likely because of increased water transport from ocean surface that is subsequently advected over land - Total water content increases over water within PBL - Total water content decreases over land within PBL - Temperature generally increases with increasing alp1 or alp2+alp3 However, unrealistic alp values had to be used to change coastal clouds How can we make the water results (e.g. more water transport, increasing cloud water content, and reducing cloud dissipation rates) occur over land? - Is ocean vertical transport of water sustaining the clouds? - Heating profiles over water are much smaller as well - Land Surface Model? 25
Land surface models RUC LSM RAP using standard PBL parameters and RUC LSM (dashed line) - MYNN scheme is optimized to use RUC LSM - RAP+RUC LSM persists clouds for much longer RUC - NOAH 26
Heat exchange coefficients are similar - RAP+RUC is more efficient at transferring heat HFX QFX FLHC FLQC q HFX FLHC Water exchange coefficients are much greater in RAP+RUC LSM - Greatest near coast - Difference over land only 27
Identification of Multi-Model Ensemble Members - WRF v3.4.1 - MODIS 20 land use categories - 12, 4 and 1.3 km horizontal resolution - 75 vertical levels - 48 ensemble members - 24 hour spin up Global Horizontal Irradiance SolarAnywhere GOES derived (visible channel) 0.01 o horizontal resolution 30 minutes temporal resolution Agglomerative Hierarchical clustering marine layer (thick clouds) x y = x i y 2 Metric Euclidean distance i N =1 Linkage Complete (max distance between pairs) Clusters number Reducing the SD for each cluster 28
Relative Euclidean Distance Identification of Multi-Model Ensemble Members Cumulus = Simplified Arakawa Schubert (SAS) Cumulus = Kain - Fritsch Cumulus scheme is most important distinction. 29
GHI Bias GHI Bias Mean Evolution of 2 and 8 Cluster Members 2 Members 8 Members Kain - Fritsch SAS Hour [PST] Hour [PST] New SAS cumulus scheme provides more clouds Unbiased mornings Too cloudy afternoons Largest bias around 1000 PST for the Kain Fritsch cumulus option around 1400 PST for the new SAS cumulus option 30
Relevance analysis of WRF outputs for postprocessing Automatic Relevance Determination (ARD) based on ANNs. Hyper-parameters control the weights distribution associated with each input. High are associated with low relevance inputs.
Data WRF runs for 4-1-2013 to 4-5-2013 using member P1M10C14S4. Target: 1 hour moving average GHI. Measured GHI for coastal station. 32
Preliminary Results A single member of the ensemble is analyzed. Most relevant WRF outputs obtained include: Shortwave downward flux (GHI, as expected) Clear sky shortwave downward flux Soil temperature Ground heat flux Accumulated total cumulus precipitation Other instantaneous and accumulated shortwave fluxes. 33 33
Post-processing of WRF output (1 ensemble) Persistence GHI (t) = GHI (t -24h) Equal weights method Ensemble average Euclidean based weights method weights proportional to the inverse of the euclidean distance Gradient descent method weights selection based on the minimization of the RMSD Analog method Analogs selected based on several meteorological variables (ghi, psfc, hgt850, hgt700, pblh, invmag, t2, RH2, windmag, and winddir) Artificial neural networks (Optimized by Genetic Algorithm) 1 hidden layer with 3 neurons, 100 chromosomes with 5000 generations. 34
Postprocessing Results 15 training days CBD is a coastal station (0.7 mi from the shore) All methods improve the overall performance of each individual ensemble member ANN GA and analogues method present the best performance, similar to the persistence. 35
WRF Conclusions WRF with cloud assimilation is much less biased than the NAM and significantly more accurate - Predict ML frequency accurately (whereas NAM rarely predicts ML) - ML evening reformation is accurate Two primary error regimes remain: - Days in which the ML dissipates too quickly (despite assimilation of clouds) - Trough days in which clouds are high (and cloud assimilation assumptions are invalid) Setting up marine layer solar forecast ensemble for SDG&E 36
Sky Imager Forecasting
Advanced Sky Imagers Hardware Development 2 years development 1 year of testing 4 deployments ready for commercialization
High Dynamic Range Imaging
Cloud Tracking raw image RBR RBR - CSL binary image cloud motion ray tracing of direct solar beam sky imager cloudmap Sky Imager Clouds detected using red-blue ratio (RBR) Clear sky library (CSL) provides detection reference Binary image projected to sky (cloudmap) Ray tracing provides ground shadows forecast domain