Forecasting the Marine Layer Front through Physical Inputs and Machine Learning Techniques

Southern California features one of the highest rates of renewable energy electric grid penetration in the US (about 14.6%). As proposed in the renewable portfolio standard of the State of California, 33% of energy consumed must be supplied by renewables by 2020. Solar energy will play a significant role in reaching this goal. Due to the intermittent nature of solar energy and its sensitivity to weather conditions, high-fidelity forecasting of solar irradiance on multiple time horizons is an enabling technology to achieve reliable and economical integration of solar power into the power grid.

The weather in the coastal area of Southern California is highly influenced by the occurrence of low marine layer clouds during the summer season. An extensive marine layer system forms over the Pacific Ocean and is frequently propelled inland, causing overcast meteorological conditions on distributed and utility scale solar farms, which in turn cause ramping events and fluctuations in power amplitude and frequency.

The formation, movement and behavior of marine layer clouds is modeled as a multivariable system that is governed by factors such as local topography, pressure gradients, coastal winds, solar radiation, etc. Multiple influences on the development of the Marine Layer Clouds have been investigated and are discussed in this work (e.g. atmospheric pressure, subsidence inversion height and strength, temperature gradients, etc.). We identify the basic parameters that can be measured or forecasted and use them as inputs for implementation in a basic operational forecast system for the marine layer. This system will cover day-ahead forecasts as well as intra-day forecasts reaching from 30 minutes to 4 hours ahead.

To increase the overall accuracy of the forecast, these inputs are combined with different machine learning techniques to form a hybrid forecasting algorithm. A database containing 10 years of binary cloud information was created using GOES-West satellite images. This database contains 15000 images and is used for pattern recognition and classification, using the k-Nearest Neighbour (kNN) algorithm to build the forecast. Identification methods for changing patterns between the first two post-sunrise images are also explored.

Results indicate that a hybrid approach is effective in forecasting marine layer clouds in Southern California, and that this hybrid approach holds promise in the development of an operational solar forecasting system for the region.