Objective and background:
Solar energy is a promising renewable energy source, notably for Singapore where the Energy Market Authority aims to increase the photovoltaic installed capacity in Singapore from 100 MWp in 2017 to 1 GWp beyond 2020. To accompany this 10-fold growth, an improvement in the forecasting of photovoltaic power is needed to cope with its meteorological variability. In particular, accurate intra-day forecasts of solar irradiance are critical.
Models based on ground measurements and advanced machine learning techniques have shown promising results for hour-ahead forecasting of solar irradiance in Singapore1. However, the accuracy of this type of models decreases with the forecasting horizon; Lauret et al.2 showed that incorporating irradiance forecast from Numerical Weather Prediction models (NWP) in addition to ground measurements in the set of predictors can improve the intra-day performances.
In this work, we propose an intra-day solar forecasting model that combines multiple outputs of the Weather Research and Forecasting mesoscale model (WRF) with ground-based measurements as predictors of a machine learning approach based on random forest.
The Solar Energy Research Institute of Singapore (SERIS) operates a network of meteorological stations across the island; eight of them are equipped with a pyranometer that measures the global horizontal irradiance (GHI) at 1-min resolution. In this work, hourly-average GHI from these stations is used over the time period from January 2014 to December 2016.
A high-resolution implementation of WRF is run for every day of the same 3 years period. The model is initiated at 8 PM SGT with a horizon of 24 hours. WRF output variables are logged hourly.
Our objective is to forecast the hourly GHI up to 7 hours ahead for one particular station. To that end, we implement a random forest, RF_wrf_multi, that takes the following groups of features as input:
To contrast the role of WRF features in the performance of the forecasts, we implement two additional random forests: RF_wrf_single that only uses features from groups 1 and 2, and RF_no_wrf that only uses features from group 1. All 3 models are trained and tuned using the years 2014 and 2015, while the year 2016 is reserved for testing.
The Mean Absolute Error (MAE) for each horizon is shown for all 3 models in the figure below. Both models using WRF features outperform RF_no_wrf for horizons beyond 2h. Furthermore, for horizons greater than 5h, the MAE of RF_wrf_multi is lower than that of RF_wrf_single.
Preliminary results showed that incorporating NWP inputs improves intra-day forecasting. Furthermore, we found that adding WRF features not directly related to the irradiance further enhances the performance of the model, in particular for larger forecasting horizons.