Short term solar irradiance time-series forecasting with machine learning

Abstract

The intermittent nature of solar irradiance caused by clouds results in rapid short term fluctuations in the power output of photovoltaics(PV)-systems. These fluctuations make this data source difficult to integrate with the grid. This study focuses on short term (0-20 minutes) forecasting of global horizon irradiance (GHI) using all sky imager (ASI) images and numerical data as input. This numerical data is locally obtained (Almeria, Spain) from June to December 2019. Installed sensors measure humidity, temperature and GHI every 15 seconds. Multiple approaches and models are proposed and compared with each other. We track clouds with optical-flow and generate a future representation of the sky. This future image is the input for a Convolutional neural network (CNN), which is trained to predict GHI. Secondly, extracted information from images and additional numerical data are plugged into different classifier models, including random forests (RF), multi-layer perceptrons (MLP), and long short-term memory networks (LSTM). We tried multiple subsets of features and multiple sequence lengths to find a good model. We categorized the features in 3 subsets and a combination of these are used as input. Models are trained to predict GHI from 0 to 20 minutes ahead, with a step size of 1 minute. We tried improving predictions by making the models predict the clear sky index (CSI), where the GHI can be derived using the Perez conversion model. We implemented this approach into the models, there was no improvement. In solar forecasting the (smart-)persistence model is a common baseline. Persistence forecasts the last observed GHI. Smart-persistence calculates the future direct normal irradiance (DNI) and combines this with the last observed CSI to predict the expected GHI. We categorize days in cloudy (C S I < 0.25), partially cloudy (C S I > 0.25,C S I < 0.75) and sunny (C S I > 0.75). We implement multiple models and show that LSTM performs best, which agrees with it’s know advantage of dealing with sequence problems. We learn what features are valuable and the amount of minutes prior prediction moment that data has predictive value. The best model(s) are able to predict statistically significantly better than the baseline for sunny and cloudy weather.