A moment in the sun

Abstract

Solar energy is now the cheapest form of electricity in history. Unfortunately, signicantly increasing the electric grid's fraction of solar energy remains challenging due to its variability, which makes balancing electricity's supply and demand more dicult. While thermal generators' ramp rate-the maximum rate at which they can change their energy generation-is nite, solar energy's ramp rate is essentially innite. Thus, accurate near-term solar forecasting, or nowcasting, is important to provide advance warnings to adjust thermal generator output in response to variations in solar generation to ensure a balanced supply and demand. To address the problem, this paper develops a general model for solar nowcasting from abundant and readily available multispectral satellite data using self-supervised learning.Specically, we develop deep auto-regressive models using convolutional neural networks (CNN) and long short-term memory networks (LSTM) that are globally trained across multiple locations to predict raw future observations of the spatio-temporal spectral data collected by the recently launched GOES-R series of satellites. Our model estimates a location's near-term future solar irradiance based on satellite observations, which we feed to a regression model trained on smaller site-specic solar data to provide near-term solar photovoltaic (PV) forecasts that account for site-specic characteristics. We evaluate our approach for dierent coverage areas and forecast horizons across 25 solar sites and show that it yields errors close to that of a model using ground-truth observations. CCS CONCEPTS• Computing methodologies ! Machine learning approaches.

Leveraging spatiotemporal information in meteorological image sequences: From feature engineering to neural networks

“…Multi-spectral context is also relevant in this application, as both visible and infrared channels provide important information. Similarly, forecasting solar radiation also benefits from providing a sequence of temporal images to a neural network architecture, as shown by Bansal et al (2021Bansal et al ( , 2022. These two applications are discussed in more detail in Section 2.…”

“…See Bansal et al (2022) and Section 4.3 for more explanation. This figure is adapted and used with due permission from Bansal et al (2022). ©Author(s) of Bansal et al (2022).…”

“…Environmental science example: The solar forecasting application discussed in Section 2 and introduced in Figure 5 uses the first-space-then-time principle. Figure 11 shows that architecture in terms of the components of Figure 10 from the work by Bansal et al (2021Bansal et al ( , 2022. The architecture first uses a CNN model to extract the spatial or channel information stored in an image tile and then applies a one-layer LSTM on top of it to extract the temporal information from the time-series visible channels from the GOES-16 satellite.…”

“…This approach allows us to train NNs without requiring labels. Bansal et al (2021Bansal et al ( , 2022 a self-supervised model to nowcast solar irradiance 15 min into the future using a combination of CNN and LSTM models on the visible channels of the GOES satellite. The key idea is that the ML model trains on the entire Continguous United States (CONUS) domain and intrinsically accounts for local geographical features without the need for external supervision or inputs.…”

“…Instead, ML approaches that model solar irradiance tend to perform better (Lago et al, 2018; Bansal and Irwin, 2020) in comparison to the other approaches. Bansal et al (2022) propose an approach for solar nowcasting by forecasting solar irradiance values from multispectral visible bands from GOES satellite data using the approach shown in Figure 5. This approach uses a neural network model that is a combination of a spatial extraction followed by temporal extraction ML model (center panel) that forecasts values of the satellite channels at a future time, , which are in turn used to predict solar output at a future time (right panel).…”