Search for optimal CNN architectures incorporating heterogeneous inputs for short-term solar PV forecasting

Cloud movement makes forecasting output from solar photovoltaic (PV) panels extremely challenging. This is a growing source of concern for electricity grid operators, responsible for ensuring that supply meets demand at every instant. A better solar PV forecast can realize value for commercial and industrial customers with solar assets. In this study, we aim to build a Convolutional Neural Network (CNN) based model to forecast power output from solar PV panels 15 minutes into the future. The inputs to the model are the power output history from PV panels and ground-based sky images for the past 15 minutes.

The key challenge is ensuring that due importance is given to each type of input. A stack of 16 historical sky images with 64 x 64 spatial resolution and 3 color channels contains 196,000 numbers while 16 historical power outputs are simply 16 numbers. If these heterogeneous inputs are not carefully integrated, the prediction error obtained by using two sources of inputs can be worse than using just one of them. Also, with deep learning architectures, there are many independent variables to tune. Limited computing time and resources prevent us from conducting an exhaustive search across all possible combinations.

In this study, we explore 28 total architectures for combining heterogeneous data inputs. We do so while devising and following a simple, yet systematic, design of experiments & analysis pipeline. Through a three-stage “funnel” approach we narrow our search to the most promising of the 28 architecture options. We find four best-in-class architectures, one for each of four major methods for heterogeneous data fusion. We test these methods on a larger dataset and find that a two-step model based on autoregressive (AR) forecasts has the best performance. This architecture has a forecast skill of 17% relative to smart persistence on the test set comprising of 9 complete sunny days and 11 complete cloudy days spread across a year. We hope our general process will be generalizable to machine learning applications beyond solar forecasting.