Machine Learning techniques to forecast energy production from Wave Energy Converters

Wave energy has recently emerged as a stronger contender when it comes to renewable energy sources. With its great potential come a few drawbacks, such as the intermittent and non-stationary nature of the sea waves, where just like other renewable energy sources their power fluctuates and makes it challenging to have it integrated into the grid. It is of great importance to predict the output power of this energy source to ensure its full integration into the smart grid. This thesis evaluates the effectiveness of employing exogenous inputs for forecasting the power output in short term horizons of 15 to 240 minutes of the Inertial Sea Wave Energy Converter (ISWEC). Wave energy is captured from the ISWEC using the inertial effects of a gyroscope. To identify the most relevant exogenous inputs for forecasting the power output, different feature selection techniques were utilized. The selected features were evaluated by feeding them to the machine learning techniques that showed the best forecasting performance in the univariate case, that are Long Short-Term Memory Encoder-Decoder Model and Long Short-Term Memory Vector-to-one Forecasting Model. Furthermore, the aggregation of the data was investigated for its behavior and efficiency, where the data with initial sampling time of 0.1s was downsampled in time steps of 1, 3,5 and 15 minutes. In particular, the effectiveness of the exogenous inputs and machine learning models for different downsamplings was evaluated. The results showed that the model’s performance improves in higher downsamplings, and that the forecasting horizon slightly increased with the inclusion of the exogenous inputs.