Solar Irradiance forecasting using Sky-Camera

Production of alternative energy sources, such as solar energy, are governed by the vagaries of weather. For instance, indoor heating demands are inversely correlated with solar radiation availability. The important question is how does one estimate solar energy production efficiently in order to solve the demand-supply mismatch in alternate energy. Weather phenomenon is a complex physical causation that are difficult to model accurately. Existing solutions to solar irradiance prediction and forecasting, are expensive (high subscription fee), require satellite readings (high computational cost and infrastructure) and are prone to high latency (two measurements per day). We propose in our research work the use of sky-camers to forecast solar irradiance.

What is a Sky-Camera?

A sky-camera is an inexpensive upward facing wide-lensed camera that can be easily deployed in solar farms and roof-tops. The high accuracy and low latency of predictions with the help of a sky-camera can give rise to many interesting applications such as demand-supply matching, energy storage optimization, and predictive panel maintenance solutions. The figure below shows an example of sky-camera deployed in the vicinity of solar farms and some sample unprocessed frames from two different sky-cameras. Feng et. al. [2] in their work talk about a similar application of short term GHI forecasting based on sky images. There have been other interesting applications as well such as the work by Shao et. al. [1] that talks about forecasting solar irradiance for use in solar race cars.

Read about other works on solar energy here

• Chasing the Sun – IBM Smarter Planet blog
• Measuring the Sun – IBM Research blog

Dataset

We introduce two publicly available datasets of sky-videos, namely Colorado [3] and Arizona [4] dataset with over a million images. Below images show the respective sky cameras installed at each of the two different locations. We outperform meteorological physics models whose parameters are tuned by coarse grained radiometric data sensed from geo-satellites for nowcasting and upto 4 hours ahead-of-time [5][6].

Proposed Approach

This research presents a deep learning approach to observe and estimate short-term weather effects from sky-videos obtained with sky-cameras and directly forecast solar irradiance. Our approach utilizes dilating convolution filters to learn a full-sky representation at varying scales via joint-training aided by auxiliary weather parameters that are sensed simultaneously. The architecture diagram for our approach is illustrated below.

Performance and Analysis

Here are two sample videos in the month of April from two datasets obtained from Arizona and Colorado in the United States respectively. We chose these videos, as they illustrate the challenges of irradiance forecasting (torrential rainy days in early summer).

Corresponding to each frame, the interpolated mean of the hypercolomns is plotted that is indicative of the focus of convolution filters. Further, the measured irradiance is plotted along with the nowcast prediction and ahead of time forecast (+1, +2, +3, & +4) hours.

The Colorado video setup (TSI) consists of a sun tracker to protect the lens and equipment from direct exposure to the sun. Hence, the autofocus allows the camera to capture the clouds more clearly. We pick a challenging video from the dataset with large intra-day variations in solar irradiance. Notice, that the error in nowcasting is higher, when the sun tracker miss-fires.

Future Work

We have extended our present work to perform future frame semantic segmentation on sky videos in order to further improve the results of our solar irradiance forecasting. Our initial results and proposed approach is available on arXiv. The below gif shows sample semantic segmentation of now predictions. The three images in the illustration are a sequence of input frames, the corresponding ground truth and semantic masks generated from our approach, respectively.

References