Today, one of the greatest technological challenges is adding renewable energies to an electric grid, with the goal being able to achieve sustainable and environmentally friendly electricity generation, which is also affordable. To make the incorporation of renewable energy sources to be successful, predictive tools and methods are required which can be used to determine in advance how much energy will be available be used in the grid to preserve the fossil fuel to fulfill the future demands. This would reduce the harmful impact on the environment and the dependence on fossil fuels.
Choosing a location for a wind farm is not an easy task. By adding deep learning techniques to energy-generation models, a team of scientists at Penn State has developed a deep learning model that can locate the best location for a wind farm and also assist with 24-hour predictions. The model is trained and validated with data from a find farm located on the island of Tenerife and show that these predictors are precise than the reference model and the current model used at the farm.
The researchers noted that the problem does not need models based on truly deep neural networks. The workflow for correctly developing, training, validating and tuning these models is enhanced by the advantages that deep learning tools can offer.
According to Guido Cervone, professor of geography, and meteorology and atmospheric science, people who plan to build a wind farm usually look for good terrain and average wind speed that is not too strong and not too weak, but consistent. Therefore, the team has found a more accurate and efficient way to look at wind predictability at specific locations, which is a key factor while building a new wind farm.
For general electrical output, location is important. However, being able to predict the amount of wind energy the farm is going to produce in 24 hours in the future is also important. Electricity suppliers purchase the energy produced by wind farms and want reliability in it. Wind farms routinely supply energy to these dealers, but they should be able to schedule 24 hours in advance about how much power they will produce.
The team, therefore, developed a deep learning algorithm called Analog Ensemble (AnEn). The method used historical sets of past observations and predictions to provide a probability model for forecast, in this case, for wind energy. The team observed that locations with higher average wind speed are related to larger degrees of forecast uncertainty which increases the difficulty to predict wind speed at these locations.
The model produces a probability result for wind production from which companies can make decisions while also understanding the risk. If the risk can be predicted, the output can also be predicted. The model is efficient as it searches for a historic forecast that matches the current given situation and provides the actual wind speeds and duration.