Abstract
One of the key applications of AI algorithms in power sector involves forecasting of stochastic renewable energy sources. To manage the generation of electricity from solar or wind effectively, accurate forecasting models are imperative. In order to achieve this goal, a sophisticated hybrid neural network formulation is discussed here in this chapter. long-short-term memory and recurrent neural networks combination is formulated for very short-term forecasting of wind speed and solar radiation. In intervals of 15 and 30 minutes, time series forecasts are made that are ahead by multiple steps. For maximum energy harvest, both point wise and probabilistic forecasting approaches are combined. Historic data is collected for solar radiation, wind speed, temperature, and relative humidity, and are used to train the model. The proposed model is compared with convolutional and LSTM neural network models individually in terms of RMSE, MAPE, MAE, and correlation, and is identified to have better forecasting accuracy.
TopIntroduction
Increasing demand of electric power across the world and the issue of global warming are having a complimenting effect, pushing the power generation trend around the globe toward environment friendly energy resources. Besides, the sustainable development goals make it imperative to develop and extract energy form renewable resources, rather than relying on the non-renewable conventional kind, whose availability and affordability are going downhill, daily. Then again, there are issues like CDM (Clean Development Mechanism), climate change, insufficient and unreliable supply of power in developing countries etc., all advising the urgency of identifying alternate and nature-friendly sources of development.
Among all the alternate power production options, nested under renewable energy, wind & solar power are the most promising substitutes. These sources being economic, as well as, being available throughout the clock (wind) (Li, Wu, & Liu, 2018), minimal maintenance requirement and ease of installation (solar), have an upper hand above the other sources. As per (Hu & Chen, 2018), wind power is also one of the most cost-effective sources, that has a huge potential to compete with the traditional fossil fuel-based power plants and is eco-friendly too. These pros have provided a rapid boost to solar & wind-based power generation throughout the world, with a growth rate of 28% per year (Varanasi & Tripathi, 2016).
Key Terms in this Chapter
Long Short-Term Memory (LSTM): LSTM is explicitly a subfield of RNN architecture, which is more stable and efficient in dealing with both long-term, as well as short-term dependency problems. It is very useful when the gap between the past and the required future values are substantial.
Deep Neural Network (DNN): A class of machine learning model. The main difference between Classical and Deep network scheme is the number of the hidden layer and the training process. Using more hidden layers, DNN can extract higher order of interrelation.
Recurrent Neural Network (RNN): RNN is a type of ANN, usually used for the forecasting of time series data. It utilizes the feedback provided by one or more units of its network as input in selecting the succeeding output. In RNN, the hidden neurons connect the hidden layer from previous time step to current time step, which is why it is called recurrent.