The novel method presented here comprises techniques for cloud coverage percentage forecasts,
cloud movement forecast and the subsequently prediction of the global horizontal irradiance
(GHI) using all-sky images and Machine Learning techniques. Such models are employed to
forecast GHI, which is necessary to make more accurate time series forecasts for photovoltaic
systems like “island solutions” for power production or for energy exchange like in virtual power
plants. All images were recorded by a hemispheric sky imager (HSI) at the Institute of Meteo rology and Climatology (IMuK) of the Leibniz University Hannover, Hannover, Germany.
This thesis is composed of three parts. First, a model to forecast the total cloud cover
five-minutes ahead by training an autoregressive neural network with Backpropagation. The
prediction results showed a reduction of both the Root Mean Square Error (RMSE) and Mean
Absolute Error (MAE) by approximately 30% compared to the reference solar persistence solar
model for various cloud conditions. Second, a model to predict the GHI up to one-hour ahead by
training a Levenberg Marquardt Backpropagation neural network. This novel method reduced
both the RMSE and the MAE of the one-hour prediction by approximately 40% under various
weather conditions. Third, for the forecasting of the cloud movement up to two-minutes ahead, a
high-resolution Deep Learning method using convolutional neural networks (CNN) was created.
By taking real cloud shapes produced by the correction of the hazy areas considering the green
signal counts pixels, predicted clouds shapes of the proposed algorithm was compared with the
persistence solar model using the Sørensen-Dice similarity coefficient (SDC). The results of the
proposed method have shown a mean SDC of 94 ± 2.6% (mean ± standard deviation) for the
first minutes outperforming the persistence solar model with a SDC of 89 ± 3.8%. Thus, the
proposed method may represent cloud shapes better than the persistence solar model. Finally,
the Bonferroni's correction was performed so that the significance level of 0.05 was corrected
to 0.05, and thus, the difference between the SDC of the proposed method and the persistence
solar model was p = 0.001 being significantly high.
The proposed methodologies may have broad application in the planning and management of
PV power production allowing more accurate forecasts of the GHI minutes ahead by targeting
primary and secondary energy control reserve.
|