Abstract
Due to the terrible impact of climate change caused by the fossil energy consumed in
the past years, countries all over the world recognized that developing renewable energy
and decreasing carbon emissions is the most important content in the power energy
revolution. Under this background, increasing the proportion of renewable energy is the
key path to achieve the target of a carbon neutral and carbon free future.(Van der Ploeg
and Rezai, 2020)
The exploration of the use of renewable energy has never stopped, however, the
difficulties are also obvious, that is, the instability of the renewable resources (except
hydroelectric power). For one thing, take solar power generation as an example, the
solar power generation mainly depends on the Global Horizontal Irradiance, which is
greatly affected by meteorological conditions, such as rainfall, humidity, wind, etc.,
which are obviously unpredictable and have randomness factors. Another reason, from
the perspective of power consumption, whether industry, commercial or residential,
almost all the power consumed during these activities is also random. From the
viewpoint of the current power grid, the power consumption and power generation must
be equal in real time, but the factors on both sides produce random factors. In order to
overcome the randomness and make the grid stable, there must be some adjustable and
stable power supply added into the grid. The main such adjustable power sources are
gas / coal / petroleum peaker, which can be dispatched and switched very fast, however,
the disadvantage is that most of the fuel is fossil fuel. In order to decrease the fossil fuel
dependence in power generation as the adjustable back up power, more accuracy in
renewable power forecasting is necessary for the feed into the power grid dispatch
centre. In this way, less fossil power would be dispatched whilst optimising the
renewable power are used.
In this study, we apply the optimized artificial neural network as the basic approach,
especially the back-propagation neural network, which is good at abstracting the nonliner
principle between multiple power generation factors and solar power output.
Considering the decay during the long-term operation of the solar plant, we involve the
UV index into the model, which introduces the effect of the long-term decay and real
time ultra-volet irradiation into the system. With more than 15 years meteorological
raw data collected, a Principal Component Analysis based pre-process is applied to the
dataset, which refine the raw data and abstract the most corresponding characteristic.
To promote the model with much more accuracy and efficiency, a novel optimize
method is applied to the model, which introduces the adaptive method through a
multiple momentum algorithm and decay learning rate. To evaluate the forecasting, the
Accuracy, MSE, MAE are used to test the forecasting result.
With all these methods and the algorithm applied, the accuracy of solar power
generation forecasting has greatly been improved, meanwhile, the training time is also
saved a lot, with the whole forecasting efficiency significantly promoted.
Key words: Renewable, Solar Power, forecast, ANN, PCA, UV index.