Sky Imaging Network for Intra-Hour Spatial Solar Forecasts

PROBLEM STATEMENT AND MOTIVATION

Power utilities are experiencing unprecedented growth of distributed solar energy. The growth in Photovoltaic (PV) installations introduces significant variability to the electric grid due to the variable nature of solar irradiance. If short-term (particularly intra-hour horizons) solar variability cannot be predicted or reduced, the integration of solar power will remain costly because of the investment on energy storage or regulation capacity, particularly for micro-grids and distribution feeders.
Solar forecasts are widely recognized as a low-cost method to mitigate the variability of solar resources.
Intra-hour forecasts of solar power are essential for power plant operating, grid balancing, real-time unit dispatching, storage system optimizing, automatic generation controlling, and operating of regulation reserves. However, most of the proposed forecasting methods are developed for a single/discrete locations but utility operators need to forecast solar generations over large areas (e.g. cities) to estimate the amount of underlying demand and the penetration of solar power. Consequently, spatial forecasts of solar irradiance are essential for planning, integrating, regulating, and managing solar power generations.

METHODOLOGY AND RESULTS

In this work, we use a network of seven hemispheric fisheye sky-imaging cameras installed and operating in the Los Angeles Basin area. This unique network of sky imaging cameras provides wide coverage of this urban region and is used to obtain high resolution spatio-temporal solar monitoring and solar forecast at low-cost. Off-the-shelf fish-eye cameras are equipped with CCD sensors that capture sky images in three 8-bit channels (Red, Green and Blue). By comparing average pixel intensity and solar irradiance obtained independently we observe that color intensity is highly correlated to the solar irradiance values. We explore this relationship to derive DHI and DNI values from the images. We don’t limit this analysis to color intensity, we expand it to six basic image features (red (R), green (G), blue (B), hue (V), saturation (S), intensity (I)) and combinations thereof. From this analysis we obtained an empirical formula to compute DHI from sky-images (corr. coef. = 0.94, R² = 0.88). Figure 1A shows the good accuracy of the DHI model.

We also develop an empirical model for estimating DNI from the sky images. This is a more complex model that requires estimating DNI’s clear-sky index depending on the cloud cover (sun not blocked by clouds k_DNI=1, sun blocked by thick clouds k_DNI=0, and sun blocked by thin clouds 0<k_DNI<1). Once DHI and DNI are estimated we obtain GHI using the trigonometric relationship between these variables. Results for the GHI Image-to-Irradiance (I2I) model are shown in Figure 1B. This model is developed using images and irradiance measurements captured in a 20-day period. The I2I model is then applied to the seven locations targeted in this work (Figure 1D) where irradiance sensors are not available.

Using the GHI data derived with the I2I model and several numerical features from the sky images we train Artificial Neural Networks (ANN) to predict 1 minute averaged GHI 5, 10 and 15 minutes ahead of time. The ANN model is then applied to an independent testing set and benchmarked against the clear-sky persistence model. Error metrics from a few locations are shown in Figure 1C. The results show that the ANN model with exogenous image features can obtain appreciable forecasting skills for these time horizons.

In the last part of this work we explore how to obtain spatial GHI forecasts from the network of sky images. Ideally high-resolution gridded solar forecasts could be obtained with a dense network of sensors.
In practice this is unattainable due to high cost and maintenance requirements.
Therefore, we explore interpolation techniques to estimate irradiance in a large geographical area. Specifically we explore inverse distance weighting (IDW) and kriging. The models were validated using the cross validation method, in which the GHI for one of the stations is held out as validation set and compared against the GHI interpolation from the other six stations. Results from the validation study indicates that the inverse distance weighting obtains better results than kriging.

FIGURE CAPTION

Figure 1: (A) DHI obtained with the I2I model compared against measured DHI. (B) Same as (A) but for GHI. (C) Error metrics for three of the locations in this study. The three columns under the model header (Persistence or ANN) correspond to the three forecast horizons (5, 10, and 15 minutes). The skill of the ANN model is relative to the rRMSE of the persistence model. (D) The map shows the location of the seven stations in the LA area. The overlaid color map shows the spatial GHI forecast obtained with the IDW interpolation of the seven stations.