Solar Irradiance Forecasting Using Multi-layer Cloud Tracking and Numerical Weather Prediction Jin Xu, Shinjae Yoo, Dantong Yu, Dong Huang, John Heiser, Paul Kalb
Solar Energy Abundant, clean, and secure renewable resource Intermittent nature of solar irradiance challenges the regulation and maintenance of grids Clouds cause fluctuations in the photovoltaic (PV) output 2
Motivation: Climatology on the East Coast 3
Motivation: Climatology on the East Coast Reliable 1-15 minute short term irradiance forecasting is essential for grid operators to control ramp events. 3
Short-term Forecasting Workflow Cloud Tracking Feature Engineering Prediction Image preprocessing Cloud cover classification Multi-layer cloud detection Segment identification Cloud tracking Data calibration Feature extraction Correlation analysis with irradiance Feature selection Linear & non-linear modeling Model and feature set evaluation and cross validation 8
Sky Imaging Based Prediction Short term ground-based sensor data collection High spatial and temporal resolution Able to capture local cloud movement TSI (Total Sky Imager) Pyranometer (irradiance sensor) TSI instrument Sample TSI image 4
Solar Irradiance Fluctuations TSI image sequence Ground irradiance 5
Limitations of Current Methods For Sky Image Processing Thin, high altitude clouds are difficult to detect Unable to differentiate between multiple cloud layers Block and pixel based tracking are noise prone Accurate forecasts are limited to 5 minutes 6
Novel Solar Irradiance Forecasting Framework Multi-layer cloud detection and tracking Cloud type classification Multi-layer detection Segment identification Numerical Weather Prediction (NWP) incorporation Extend the forecasting horizon to 15 minutes 7
Novel Cloud Tracking Pipeline 9
Novel Cloud Tracking Pipeline Original image 9
Novel Cloud Tracking Pipeline Image undistortion Shadowband dispatch 10
Novel Cloud Tracking Pipeline Cloud Cover 11
Novel Cloud Tracking Pipeline 12
Novel Cloud Tracking Pipeline 12
Novel Cloud Tracking Pipeline Cloud segmentation 13
Novel Cloud Tracking Pipeline Normalized Cross Correlation Algorithm 17
Novel Cloud Tracking Pipeline Example cloud segments at time t-1 and t and corresponding predicted motion vectors t-1 t 14
Method Comparison Segment based cloud tracking with multi-layer detection Block based tracking 15
Multi-layer Detection and Segment Identification 20
Feature Engineering: TSI Image Features 16
Feature Engineering: NWP Features Category 1 Rain 2 Light Rain 3 Overcast 4 Mostly Cloudy 5 Partly Cloudy 6 Scattered Cloud 7 Fog 8 Clear 9 Sunny 17
Full Feature Set TSI Feature Set circumsolar RBR, motion vector length/count/sum, cloud cover mean/variance, blue channel max/min, shadowband brightness. Weather Feature Set category, temperature, humidity, pressure, wind direction, and wind speed. 22
Forecast Models Persistent Model (PM) Linear Regression (LR) One TSI feature (LR_ RBR,GHI ) All features (LR_ all ) Support Vector Regression (SVR) Linear kernel (SVR lnr _all) Radial Basis Function kernel (SVR RBF _all) 18
Forecast Models Persistent Model (PM) Linear Regression (LR) with only one TSI feature (LR_RBR,GHI) with all features (LR_all) 23
Forecast Models Support Vector Regression (SVR) using linear kernel (SVR_lnr) using Radial Basis Function kernel (SVR_RBF) 24
1 to 15 Minute Forecasting Using Different Models 19
1 to 15 Minute Forecasting Using Different Feature Sets 20
1 to 15 Minute Forecasting Using Different Cloud Tracking methods 29
Example Forecasts on a Cloudy Day Forecasting of 1-min averaged GHI for 1, 5, 10 and 15 min ahead using persistent model and SVR-RBF, between 10:00 am and 14:00 pm on June 6th, 2013. 30
Summary Novel TSI image processing pipeline Differentiate low, thick clouds from high, thin clouds Tracking a complete cloud Significant improvement on short term (1-15 minute) solar irradiance forecasting model Incorporate TSI image features & NWP features Average of 21% improvement over baseline model 21
Existing GHI forecast methods Statistical based models Persistent model (PM), benchmark, which directly uses the present irradiance as the prediction. ARMA, ANN, etc., use historical GHI data to train models to predict future irradiances. Physics based models Numerical Weather Prediction (NWP), utilize meteorological observations and measurements. Cloud imagery based techniques (Deterministic) -- satellite based -- ground based (Total Sky Imager, Whole Sky Imager, etc.) 32
Full Feature Set 33
Linear Chain CRF G = (V; F;E) 34
1 to 5 Minute Forecasting Using Different Models 18
1 to 5 Minute Forecasting Using Different Feature Sets
MAE for different numbers of states of CRF and HMM